Datasets:

jamescalam

/

llama-2-arxiv-papers-chunked

like 17

1411.6836

37

4.4. Convolutional layer analysis We study the performance of filter banks extracted from different layers of a CNN in the FV-CNN framework. We use the VGG-M network which has five convolutional layers. Results on various datasets, obtained as in Sect. 4.1 and 4.2, are shown in Fig. 4. In addition we also show the performance using FVs constructed from dense SIFT using a number of words such that the resulting FV is roughly the same size of FV-CNN. The CNN filter banks from layer 3 and beyond significantly outperform SIFT. The performance monotonically improves from layer one to five. 5. Conclusions We have conducted a range of experiments on material and texture attribute recognition in a large dataset of textures in clutter. This benchmark was derived from OpenSurfaces, an earlier contribution of the computer graphics community, highlights the potential for collaboration between computer graphics and vision communities. We have also evaluated a number of state-of-the-art texture descriptors on these and many other benchmarks. Our main finding is that orderless pooling of convolutional neural network features is a remarkably good texture descriptor, versatile enough to dubbed as a scene and object descriptor, resulting in new

1411.6836

Deep convolutional filter banks for texture recognition and segmentation

Research in texture recognition often concentrates on the problem of material recognition in uncluttered conditions, an assumption rarely met by applications. In this work we conduct a first study of material and describable texture at- tributes recognition in clutter, using a new dataset derived from the OpenSurface texture repository. Motivated by the challenge posed by this problem, we propose a new texture descriptor, D-CNN, obtained by Fisher Vector pooling of a Convolutional Neural Network (CNN) filter bank. D-CNN substantially improves the state-of-the-art in texture, mate- rial and scene recognition. Our approach achieves 82.3% accuracy on Flickr material dataset and 81.1% accuracy on MIT indoor scenes, providing absolute gains of more than 10% over existing approaches. D-CNN easily trans- fers across domains without requiring feature adaptation as for methods that build on the fully-connected layers of CNNs. Furthermore, D-CNN can seamlessly incorporate multi-scale information and describe regions of arbitrary shapes and sizes. Our approach is particularly suited at lo- calizing stuff categories and obtains state-of-the-art re- sults on MSRC segmentation dataset, as well as promising results on recognizing materials and surface attributes in clutter on the OpenSurfaces dataset.

http://arxiv.org/pdf/1411.6836

Accepted to CVPR15

cs.CV

20141125

20150709

1411.6836

38

orderless pooling of convolutional neural network features is a remarkably good texture descriptor, versatile enough to dubbed as a scene and object descriptor, resulting in new state-of-the-art performance in several benchmarks. brick cardboard carpet_rug ceramic concrete fabric_cloth food glass granite_marble hair other laminate leather metal painted paper_tissue plastic_clear plastic_opaque stone tile wallpaper wood(a) (b) (c) (d) (e) (f) Figure 2. OS material recognition results. Example test image with material recognition and segmentation on the OS dataset. (a) original image. (b) ground truth segmentations from the OpenSurfaces repository (note that not all pixels are annotated). (c) FC-CNN and crispregion proposals segmentation results. (d) incorrectly predicted pixels (restricted to the ones annotated). (e-f) the same, but for FV-CNN. building grass tree cow sheep sky aeroplane water face car bicycle flower sign bird book chair road cat dog body boat (a) (b) (c) (d) (e) (f) Figure 3. MSRC object segmentation results. (a) image, (b) ground-truth, (c-d) FC-CNN, (d-e) FV-CNN segmentation and errors.

1411.6836

Deep convolutional filter banks for texture recognition and segmentation

Research in texture recognition often concentrates on the problem of material recognition in uncluttered conditions, an assumption rarely met by applications. In this work we conduct a first study of material and describable texture at- tributes recognition in clutter, using a new dataset derived from the OpenSurface texture repository. Motivated by the challenge posed by this problem, we propose a new texture descriptor, D-CNN, obtained by Fisher Vector pooling of a Convolutional Neural Network (CNN) filter bank. D-CNN substantially improves the state-of-the-art in texture, mate- rial and scene recognition. Our approach achieves 82.3% accuracy on Flickr material dataset and 81.1% accuracy on MIT indoor scenes, providing absolute gains of more than 10% over existing approaches. D-CNN easily trans- fers across domains without requiring feature adaptation as for methods that build on the fully-connected layers of CNNs. Furthermore, D-CNN can seamlessly incorporate multi-scale information and describe regions of arbitrary shapes and sizes. Our approach is particularly suited at lo- calizing stuff categories and obtains state-of-the-art re- sults on MSRC segmentation dataset, as well as promising results on recognizing materials and surface attributes in clutter on the OpenSurfaces dataset.

http://arxiv.org/pdf/1411.6836

Accepted to CVPR15

cs.CV

20141125

20150709

1411.6836

39

Acknowledgements M. Cimpoi was supported by ERC grant VisRec no. 228180 and the XRCE UAC grant. S. Maji acknowledges a faculty startup grant from UMass, Amherst.References [1] E. H. Adelson. On seeing stuff: The perception of materials by humans and machines. SPIE , 4299, 2001. 1 [2] M. Amadasun and R. King. Textural features corresponding to textural properties. Systems, Man, and Cybernetics , 19(5), 1989. 1 [3] P. Arbel ´aez, J. Pont-Tuset, J. T. Barron, F. Marques, and J. Malik. Multiscale combinatorial grouping. CVPR, 2014. 4, 5, 7 [4] S. Bell, P. Upchurch, N. Snavely, and K. Bala. Opensurfaces: A richly annotated catalog of surface appearance. In Proc. SIGGRAPH , 2013. 2 [5] B. Caputo, E. Hayman, and P. Mallikarjuna. Class-specific material categorisation. In ICCV , 2005. 1, 4

1411.6836

Deep convolutional filter banks for texture recognition and segmentation

Research in texture recognition often concentrates on the problem of material recognition in uncluttered conditions, an assumption rarely met by applications. In this work we conduct a first study of material and describable texture at- tributes recognition in clutter, using a new dataset derived from the OpenSurface texture repository. Motivated by the challenge posed by this problem, we propose a new texture descriptor, D-CNN, obtained by Fisher Vector pooling of a Convolutional Neural Network (CNN) filter bank. D-CNN substantially improves the state-of-the-art in texture, mate- rial and scene recognition. Our approach achieves 82.3% accuracy on Flickr material dataset and 81.1% accuracy on MIT indoor scenes, providing absolute gains of more than 10% over existing approaches. D-CNN easily trans- fers across domains without requiring feature adaptation as for methods that build on the fully-connected layers of CNNs. Furthermore, D-CNN can seamlessly incorporate multi-scale information and describe regions of arbitrary shapes and sizes. Our approach is particularly suited at lo- calizing stuff categories and obtains state-of-the-art re- sults on MSRC segmentation dataset, as well as promising results on recognizing materials and surface attributes in clutter on the OpenSurfaces dataset.

http://arxiv.org/pdf/1411.6836

Accepted to CVPR15

cs.CV

20141125

20150709

1411.6836

40

categorisation. In ICCV , 2005. 1, 4 [6] K. Chatfield, K. Simonyan, A. Vedaldi, and A. Zisserman. Return of the devil in the details: Delving deep into convolutional nets. In Proc. BMVC , 2014. 2, 3, 4, 6 [7] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, , and A. Vedaldi. Describing textures in the wild. In Proc. CVPR , 2014. 1, 3, 4, 5 [8] K. J. Dana, B. van Ginneken, S. K. Nayar, and J. J. Koenderink. Reflectance and texture of real world surfaces. ACM Transactions on Graphics , 18(1):1–34, 1999. 1 [9] A. Efros and T. Leung. Texture synthesis by non-parametric sampling. In CVPR , volume 2, 1999. 1 [10] M. Everingham, A. Zisserman, C. Williams, and L. V . Gool. The PASCAL visual obiect classes challenge 2007 (VOC2007) results. Technical report, Pascal Challenge, 2007. 3, 4

1411.6836

Deep convolutional filter banks for texture recognition and segmentation

Research in texture recognition often concentrates on the problem of material recognition in uncluttered conditions, an assumption rarely met by applications. In this work we conduct a first study of material and describable texture at- tributes recognition in clutter, using a new dataset derived from the OpenSurface texture repository. Motivated by the challenge posed by this problem, we propose a new texture descriptor, D-CNN, obtained by Fisher Vector pooling of a Convolutional Neural Network (CNN) filter bank. D-CNN substantially improves the state-of-the-art in texture, mate- rial and scene recognition. Our approach achieves 82.3% accuracy on Flickr material dataset and 81.1% accuracy on MIT indoor scenes, providing absolute gains of more than 10% over existing approaches. D-CNN easily trans- fers across domains without requiring feature adaptation as for methods that build on the fully-connected layers of CNNs. Furthermore, D-CNN can seamlessly incorporate multi-scale information and describe regions of arbitrary shapes and sizes. Our approach is particularly suited at lo- calizing stuff categories and obtains state-of-the-art re- sults on MSRC segmentation dataset, as well as promising results on recognizing materials and surface attributes in clutter on the OpenSurfaces dataset.

http://arxiv.org/pdf/1411.6836

Accepted to CVPR15

cs.CV

20141125

20150709

1411.6836

41

PASCAL visual obiect classes challenge 2007 (VOC2007) results. Technical report, Pascal Challenge, 2007. 3, 4 [11] V . Ferrari and A. Zisserman. Learning visual attributes. In Proc. NIPS , 2007. 1 [12] D. Forsyth. Shape from texture and integrability. In ICCV , volume 2, pages 447–452. IEEE, 2001. 1 [13] J. G ˚arding. Shape from texture for smooth curved surfaces in perspective projection. Journal of Mathematical Imaging and Vision , 2(4):327–350, 1992. 1 [14] R. B. Girshick, J. Donahue, T. Darrell, and J. Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. InProc. CVPR , 2014. 2, 3 [15] Y . Gong, L. Wang, R. Guo, and S. Lazebnik. Multi-scale orderless pooling of deep convolutional activation features. In Proc. ECCV , 2014. 2, 3, 6 [16] B. Hariharan, P. Arbel ´aez, R. Girshick, and J. Malik. Simultaneous detection and segmentation. In Computer Vision–ECCV 2014 , pages

1411.6836

Deep convolutional filter banks for texture recognition and segmentation

Research in texture recognition often concentrates on the problem of material recognition in uncluttered conditions, an assumption rarely met by applications. In this work we conduct a first study of material and describable texture at- tributes recognition in clutter, using a new dataset derived from the OpenSurface texture repository. Motivated by the challenge posed by this problem, we propose a new texture descriptor, D-CNN, obtained by Fisher Vector pooling of a Convolutional Neural Network (CNN) filter bank. D-CNN substantially improves the state-of-the-art in texture, mate- rial and scene recognition. Our approach achieves 82.3% accuracy on Flickr material dataset and 81.1% accuracy on MIT indoor scenes, providing absolute gains of more than 10% over existing approaches. D-CNN easily trans- fers across domains without requiring feature adaptation as for methods that build on the fully-connected layers of CNNs. Furthermore, D-CNN can seamlessly incorporate multi-scale information and describe regions of arbitrary shapes and sizes. Our approach is particularly suited at lo- calizing stuff categories and obtains state-of-the-art re- sults on MSRC segmentation dataset, as well as promising results on recognizing materials and surface attributes in clutter on the OpenSurfaces dataset.

http://arxiv.org/pdf/1411.6836

Accepted to CVPR15

cs.CV

20141125

20150709

1411.6836

42

detection and segmentation. In Computer Vision–ECCV 2014 , pages 297–312. Springer, 2014. 2, 7 [17] E. Hayman, B. Caputo, M. Fritz, and J.-O. Eklundh. On the significance of real-world conditions for material classification. ECCV , 2004. 4 [18] K. He, X. Zhang, S. Ren, and J. Sun. Spatial pyramid pooling in deep convolutional networks for visual recognition. In Proc. ECCV , 2014. 2 [19] P. Isola, D. Zoran, D. Krishnan, and E. H. Adelson. Crisp boundary detection using pointwise mutual information. In Proc. ECCV , 2014. 4, 7 [20] A. Jain and F. Farrokhnia. Unsupervised texture segmentation using gabor filters. Pattern recognition , 24(12):1167–1186, 1991. 1 [21] Y . Jia. Caffe: An open source convolutional architecture for fast feature embedding. http://caffe.berkeleyvision.org/ , 2013. 3, 4, 6 [22] B. Julesz and J. R. Bergen. Textons, the fundamental elements in

1411.6836

Deep convolutional filter banks for texture recognition and segmentation

Research in texture recognition often concentrates on the problem of material recognition in uncluttered conditions, an assumption rarely met by applications. In this work we conduct a first study of material and describable texture at- tributes recognition in clutter, using a new dataset derived from the OpenSurface texture repository. Motivated by the challenge posed by this problem, we propose a new texture descriptor, D-CNN, obtained by Fisher Vector pooling of a Convolutional Neural Network (CNN) filter bank. D-CNN substantially improves the state-of-the-art in texture, mate- rial and scene recognition. Our approach achieves 82.3% accuracy on Flickr material dataset and 81.1% accuracy on MIT indoor scenes, providing absolute gains of more than 10% over existing approaches. D-CNN easily trans- fers across domains without requiring feature adaptation as for methods that build on the fully-connected layers of CNNs. Furthermore, D-CNN can seamlessly incorporate multi-scale information and describe regions of arbitrary shapes and sizes. Our approach is particularly suited at lo- calizing stuff categories and obtains state-of-the-art re- sults on MSRC segmentation dataset, as well as promising results on recognizing materials and surface attributes in clutter on the OpenSurfaces dataset.

http://arxiv.org/pdf/1411.6836

Accepted to CVPR15

cs.CV

20141125

20150709

1411.6836

43

2013. 3, 4, 6 [22] B. Julesz and J. R. Bergen. Textons, the fundamental elements in preattentive vision and perception of textures. Bell System Technical Journal , 62(6, Pt 3):1619–1645, Jul-Aug 1983. 1 [23] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks. In Proc. NIPS , 2012. 2, 3 [24] L. Ladicky, C. Russell, P. Kohli, and P. Torr. Graph cut based inference with co-occurrence statistics. In Proc. ECCV , pages 239–253, 2010. 5 [25] L. Ladicky, P. Sturgess, K. Alahari, C. Russell, and P. Torr. What, where and how many? combining object detectors and crfs. In Proc. ECCV , 2010. 7 [26] T. Leung and J. Malik. Representing and recognizing the visual appearance of materials using three-dimensional textons. International Journal of Computer Vision , 43(1):29–44, 2001. 1

1411.6836

Deep convolutional filter banks for texture recognition and segmentation

Research in texture recognition often concentrates on the problem of material recognition in uncluttered conditions, an assumption rarely met by applications. In this work we conduct a first study of material and describable texture at- tributes recognition in clutter, using a new dataset derived from the OpenSurface texture repository. Motivated by the challenge posed by this problem, we propose a new texture descriptor, D-CNN, obtained by Fisher Vector pooling of a Convolutional Neural Network (CNN) filter bank. D-CNN substantially improves the state-of-the-art in texture, mate- rial and scene recognition. Our approach achieves 82.3% accuracy on Flickr material dataset and 81.1% accuracy on MIT indoor scenes, providing absolute gains of more than 10% over existing approaches. D-CNN easily trans- fers across domains without requiring feature adaptation as for methods that build on the fully-connected layers of CNNs. Furthermore, D-CNN can seamlessly incorporate multi-scale information and describe regions of arbitrary shapes and sizes. Our approach is particularly suited at lo- calizing stuff categories and obtains state-of-the-art re- sults on MSRC segmentation dataset, as well as promising results on recognizing materials and surface attributes in clutter on the OpenSurfaces dataset.

http://arxiv.org/pdf/1411.6836

Accepted to CVPR15

cs.CV

20141125

20150709

1411.6836

44

Journal of Computer Vision , 43(1):29–44, 2001. 1 [27] L. Liu, L. Wang, and X. Liu. In defense of soft-assignment coding. InComputer Vision (ICCV), 2011 IEEE International Conference on , pages 2486–2493. IEEE, 2011. 1[28] D. G. Lowe. Object recognition from local scale-invariant features. InProc. ICCV , 1999. 1 [29] B. Manjunath and R. Chellappa. Unsupervised texture segmentation using markov random field models. IEEE Transactions on Pattern Analysis and Machine Intelligence , 13(5):478–482, 1991. 1 [30] M. Oquab, L. Bottou, I. Laptev, and J. Sivic. Learning and Transferring Mid-Level Image Representations using Convolutional Neural Networks. In Proc. CVPR , 2014. 2 [31] O. M. Parkhi, K. Simonyan, A. Vedaldi, and A. Zisserman. A compact and discriminative face track descriptor. In Proc. CVPR , 2014. 4

1411.6836

Deep convolutional filter banks for texture recognition and segmentation

Research in texture recognition often concentrates on the problem of material recognition in uncluttered conditions, an assumption rarely met by applications. In this work we conduct a first study of material and describable texture at- tributes recognition in clutter, using a new dataset derived from the OpenSurface texture repository. Motivated by the challenge posed by this problem, we propose a new texture descriptor, D-CNN, obtained by Fisher Vector pooling of a Convolutional Neural Network (CNN) filter bank. D-CNN substantially improves the state-of-the-art in texture, mate- rial and scene recognition. Our approach achieves 82.3% accuracy on Flickr material dataset and 81.1% accuracy on MIT indoor scenes, providing absolute gains of more than 10% over existing approaches. D-CNN easily trans- fers across domains without requiring feature adaptation as for methods that build on the fully-connected layers of CNNs. Furthermore, D-CNN can seamlessly incorporate multi-scale information and describe regions of arbitrary shapes and sizes. Our approach is particularly suited at lo- calizing stuff categories and obtains state-of-the-art re- sults on MSRC segmentation dataset, as well as promising results on recognizing materials and surface attributes in clutter on the OpenSurfaces dataset.

http://arxiv.org/pdf/1411.6836

Accepted to CVPR15

cs.CV

20141125

20150709

1411.6836

45

4 [32] F. Perronnin, J. S ´anchez, and T. Mensink. Improving the fisher kernel for large-scale image classification. In Computer Vision–ECCV 2010 , pages 143–156. Springer, 2010. 1 [33] X. Qi, R. Xiao, C. G. Li, Y . Qiao, J. Guo, and X. Tang. Pairwise rotation invariant co-occurrence local binary pattern. PAMI , 36(11):2199–2213, Nov 2014. 5 [34] A. Quattoni and A. Torralba. Recognizing indoor scenes. In Proc. CVPR , 2009. 2, 4 [35] A. S. Razavin, H. Azizpour, J. Sullivan, and S. Carlsson. Cnn features off-the-shelf: An astounding baseline for recognition. In DeepVision workshop , 2014. 3 [36] G. Schwartz and K. Nishino. Visual material traits: Recognizing per-pixel material context. In Proc. CVCP , 2013. 1 [37] L. Sharan, C. Liu, R. Rosenholtz, and E. H. Adelson. Recognizing materials using perceptually inspired features. International Journal

1411.6836

Deep convolutional filter banks for texture recognition and segmentation

Research in texture recognition often concentrates on the problem of material recognition in uncluttered conditions, an assumption rarely met by applications. In this work we conduct a first study of material and describable texture at- tributes recognition in clutter, using a new dataset derived from the OpenSurface texture repository. Motivated by the challenge posed by this problem, we propose a new texture descriptor, D-CNN, obtained by Fisher Vector pooling of a Convolutional Neural Network (CNN) filter bank. D-CNN substantially improves the state-of-the-art in texture, mate- rial and scene recognition. Our approach achieves 82.3% accuracy on Flickr material dataset and 81.1% accuracy on MIT indoor scenes, providing absolute gains of more than 10% over existing approaches. D-CNN easily trans- fers across domains without requiring feature adaptation as for methods that build on the fully-connected layers of CNNs. Furthermore, D-CNN can seamlessly incorporate multi-scale information and describe regions of arbitrary shapes and sizes. Our approach is particularly suited at lo- calizing stuff categories and obtains state-of-the-art re- sults on MSRC segmentation dataset, as well as promising results on recognizing materials and surface attributes in clutter on the OpenSurfaces dataset.

http://arxiv.org/pdf/1411.6836

Accepted to CVPR15

cs.CV

20141125

20150709

1411.6836

46

[37] L. Sharan, C. Liu, R. Rosenholtz, and E. H. Adelson. Recognizing materials using perceptually inspired features. International Journal of Computer Vision , 103(3):348–371, 2013. 1, 4, 5 [38] L. Sharan, R. Rosenholtz, and E. H. Adelson. Material perceprion: What can you see in a brief glance? Journal of Vision , 9:784(8), 2009. 1 [39] K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. CoRR , abs/1409.1556, 2014. 2, 4 [40] M. Sulc and J. Matas. Fast features invariant to rotation and scale of texture. Technical report, 2014. 5 [41] M. Varma and A. Zisserman. Texture classification: Are filter banks necessary? In CVPR , volume 2, pages II–691. IEEE, 2003. 1 [42] J. Wang, J. Yang, K. Yu, F. Lv, T. Huang, and Y . Gong. Localityconstrained linear coding for image classification. In Computer Vision and Pattern Recognition (CVPR), 2010 IEEE Conference on ,

1411.6836

Deep convolutional filter banks for texture recognition and segmentation

Research in texture recognition often concentrates on the problem of material recognition in uncluttered conditions, an assumption rarely met by applications. In this work we conduct a first study of material and describable texture at- tributes recognition in clutter, using a new dataset derived from the OpenSurface texture repository. Motivated by the challenge posed by this problem, we propose a new texture descriptor, D-CNN, obtained by Fisher Vector pooling of a Convolutional Neural Network (CNN) filter bank. D-CNN substantially improves the state-of-the-art in texture, mate- rial and scene recognition. Our approach achieves 82.3% accuracy on Flickr material dataset and 81.1% accuracy on MIT indoor scenes, providing absolute gains of more than 10% over existing approaches. D-CNN easily trans- fers across domains without requiring feature adaptation as for methods that build on the fully-connected layers of CNNs. Furthermore, D-CNN can seamlessly incorporate multi-scale information and describe regions of arbitrary shapes and sizes. Our approach is particularly suited at lo- calizing stuff categories and obtains state-of-the-art re- sults on MSRC segmentation dataset, as well as promising results on recognizing materials and surface attributes in clutter on the OpenSurfaces dataset.

http://arxiv.org/pdf/1411.6836

Accepted to CVPR15

cs.CV

20141125

20150709

1411.6836

47

pages 3360–3367. IEEE, 2010. 1 [43] L. Wei and M. Levoy. Fast texture synthesis using treestructured vector quantization. In SIGGRAPH , pages 479–488. ACM Press/Addison-Wesley Publishing Co., 2000. 1 [44] Y . Wei, W. Xia, J. Huang, B. Ni, J. Dong, Y . Zhao, and S. Yan. Cnn: Single-label to multi-label. 2014. 5, 6 [45] P. Welinder, S. Branson, T. Mita, C. Wah, and F. Schroff. Caltechucsd birds 200. Technical report, Caltech-UCSD, 2010. 4 [46] N. Zhang, J. Donahue, R. Girshickr, and T. Darrell. Part-based RCNNs for fine-grained category detection. In Proc. ECCV , 2014. 5, 6 [47] B. Zhou, A. Lapedriza, J. Xiao, A. Torralba, and A. Oliva. Learning deep features for scene recognition using places database. In Proc. NIPS , 2014. 2, 5, 6, 7

1411.6836

Deep convolutional filter banks for texture recognition and segmentation

Research in texture recognition often concentrates on the problem of material recognition in uncluttered conditions, an assumption rarely met by applications. In this work we conduct a first study of material and describable texture at- tributes recognition in clutter, using a new dataset derived from the OpenSurface texture repository. Motivated by the challenge posed by this problem, we propose a new texture descriptor, D-CNN, obtained by Fisher Vector pooling of a Convolutional Neural Network (CNN) filter bank. D-CNN substantially improves the state-of-the-art in texture, mate- rial and scene recognition. Our approach achieves 82.3% accuracy on Flickr material dataset and 81.1% accuracy on MIT indoor scenes, providing absolute gains of more than 10% over existing approaches. D-CNN easily trans- fers across domains without requiring feature adaptation as for methods that build on the fully-connected layers of CNNs. Furthermore, D-CNN can seamlessly incorporate multi-scale information and describe regions of arbitrary shapes and sizes. Our approach is particularly suited at lo- calizing stuff categories and obtains state-of-the-art re- sults on MSRC segmentation dataset, as well as promising results on recognizing materials and surface attributes in clutter on the OpenSurfaces dataset.

http://arxiv.org/pdf/1411.6836

Accepted to CVPR15

cs.CV

20141125

20150709

1411.6836

48

deep features for scene recognition using places database. In Proc. NIPS , 2014. 2, 5, 6, 7 [48] X. Zhou, K. Yu, T. Zhang, and T. S. Huang. Image classification using super-vector coding of local image descriptors. In Computer Vision–ECCV 2010 , pages 141–154. Springer, 2010. 1

1411.6836

Deep convolutional filter banks for texture recognition and segmentation

Research in texture recognition often concentrates on the problem of material recognition in uncluttered conditions, an assumption rarely met by applications. In this work we conduct a first study of material and describable texture at- tributes recognition in clutter, using a new dataset derived from the OpenSurface texture repository. Motivated by the challenge posed by this problem, we propose a new texture descriptor, D-CNN, obtained by Fisher Vector pooling of a Convolutional Neural Network (CNN) filter bank. D-CNN substantially improves the state-of-the-art in texture, mate- rial and scene recognition. Our approach achieves 82.3% accuracy on Flickr material dataset and 81.1% accuracy on MIT indoor scenes, providing absolute gains of more than 10% over existing approaches. D-CNN easily trans- fers across domains without requiring feature adaptation as for methods that build on the fully-connected layers of CNNs. Furthermore, D-CNN can seamlessly incorporate multi-scale information and describe regions of arbitrary shapes and sizes. Our approach is particularly suited at lo- calizing stuff categories and obtains state-of-the-art re- sults on MSRC segmentation dataset, as well as promising results on recognizing materials and surface attributes in clutter on the OpenSurfaces dataset.

http://arxiv.org/pdf/1411.6836

Accepted to CVPR15

cs.CV

20141125

20150709

1412.0623

0

Material Recognition in the Wild with the Materials in Context Database Sean BellPaul UpchurchNoah Snavely Kavita Bala Department of Computer Science, Cornell University fsbell,paulu,snavely,kb g@cs.cornell.edu Abstract Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC ), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC , we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

1

architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation. 1. Introduction Material recognition plays a critical role in our understanding of and interactions with the world. To tell whether a surface is easy to walk on, or what kind of grip to use to pick up an object, we must recognize the materials that make up our surroundings. Automatic material recognition can be useful in a variety of applications, including robotics, product search, and image editing for interior design. But recognizing materials in real-world images is very challenging. Many categories of materials, such as fabric or wood, are visually very rich and span a diverse range of appearances. Materials can further vary in appearance due to lighting and shape. Some categories, such as plastic and ceramic, are ofAuthors contributed equally CNN Sliding CNN“wood” Transfer weights(trained) (discretely

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

2

CNN Sliding CNN“wood” Transfer weights(trained) (discretely optimized)(fixed)AMT AMT AMT MINC 3 million patches OpenSurfacesFlickr Houzz(a) Constructing MINC Material Labels(b) Patch material classification (c) Full scene material classificationP(material) Dense CRF PatchFigure 1. Overview. (a) We construct a new dataset by combining OpenSurfaces [ 1] with a novel three-stage Amazon Mechanical Turk (AMT) pipeline. (b)We train various CNNs on patches from MINC to predict material labels. (c)We transfer the weights to a fully convolutional CNN to efficiently generate a probability map across the image; we then use a fully connected CRF to predict the material at every pixel. ten smooth and featureless, requiring reasoning about subtle cues or context to differentiate between them. Large-scale datasets (e.g., ImageNet [ 21], SUN [ 31,19] and Places [ 34]) combined with convolutional neural networks (CNNs) have been key to recent breakthroughs in object recognition and scene classification. Material recognition is similarly poised for advancement through large-scale data and learning. To date, progress in material recognition

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

3

object recognition and scene classification. Material recognition is similarly poised for advancement through large-scale data and learning. To date, progress in material recognition has been facilitated by moderate-sized datasets like the Flickr Material Database (FMD) [ 26]. FMD contains ten material categories, each with 100 samples drawn from Flickr photos. These images were carefully selected to illustrate a wide range of appearances for these categories. FMD has been used in research on new features and learning methods for material perception and recognition [ 17,10,20,25]. While FMD was an important step towards material recognition, it is not sufficient for classifying materials in real-world im1arXiv:1412.0623v2 [cs.CV] 14 Apr 2015 agery. This is due to the relatively small set of categories, the relatively small number of images per category, and also because the dataset has been designed around hand-picked iconic images of materials. The OpenSurfaces dataset [ 1] addresses some of these problems by introducing 105,000 material segmentations from real-world images, and is significantly larger than FMD. However, in OpenSurfaces many material categories are under-sampled, with only tens of images.

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

4

material categories are under-sampled, with only tens of images. A major contribution of our paper is a new, well-sampled material dataset, called the Materials in Context Database (MINC ), with 3 million material samples. MINC is more diverse, has more examples in less common categories, and is much larger than existing datasets. MINC draws data from both Flickr images, which include many “regular” scenes, as well as Houzz images from professional photographers of staged interiors. These sources of images each have different characteristics that together increase the range of materials that can be recognized. See Figure 2 for examples of our data. We make our full dataset available online at http: //minc.cs.cornell.edu/. We use this data for material recognition by training different CNN architectures on this new dataset. We perform experiments that illustrate the effect of network architecture, image context, and training data size on subregions (i.e., patches) of a full scene image. Further, we build on our patch classification results and demonstrate simultaneous material recognition and segmentation of an image by performing dense classification over the image with a fully connected conditional random field (CRF) model [ 12]. By

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

5

performing dense classification over the image with a fully connected conditional random field (CRF) model [ 12]. By replacing the fully connected layers of the CNN with convolutional layers [ 24], the computational burden is significantly lower than a naive sliding window approach. In summary, we make two new contributions: We introduce a new material dataset, MINC , and 3stage crowdsourcing pipeline for efficiently collecting millions of click labels (Section 3.2). Our new semantic segmentation method combines a fully-connected CRF with unary predictions based on CNN learned features (Section 4.2) for simultaneous material recognition and segmentation. 2. Prior Work Material Databases. Much of the early work on material recognition focused on classifying specific instances of textures or material samples. For instance, the CUReT [ 4] database contains 61 material samples, each captured under 205 different lighting and viewing conditions. This led to research on the task of instance-level texture or material classification [ 15,30], and an appreciation of the challenges of building features that are invariant to pose and illumination.

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

6

building features that are invariant to pose and illumination. Later, databases with more diverse examples from each ma-terial category began to appear, such as KTH-TIPS [ 9,2], and led explorations of how to generalize from one example of a material to another—from one sample of wood to a completely different sample, for instance. Real-world texture attributes have also recently been explored [3]. In the domain of categorical material databases, Sharan et al. released FMD [ 26] (described above). Subsequently, Bell et al. released OpenSurfaces [ 1] which contains over 20,000 real-world scenes labeled with both materials and objects, using a multi-stage crowdsourcing pipeline. Because OpenSurfaces images are drawn from consumer photos on Flickr, material samples have real-world context, in contrast to prior databases (CUReT, KTH-TIPS, FMD) which feature cropped stand-alone samples. While OpenSurfaces is a good starting point for a material database, we substantially expand it with millions of new labels. Material recognition. Much prior work on material recognition has focused on the classification problem (categorizing an image patch into a set of material categories), often using

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

7

Material recognition. Much prior work on material recognition has focused on the classification problem (categorizing an image patch into a set of material categories), often using hand-designed image features. For FMD, Liu et al. [17] introduced reflectance-based edge features in conjunction with other general image features. Hu et al. [10] proposed features based on variances of oriented gradients. Qi et al. [20] introduced a pairwise local binary pattern (LBP) feature. Liet al. [16] synthesized a dataset based on KTH-TIPS2 and built a classifier from LBP and dense SIFT. Timofte et al. [29] proposed a classification framework with minimal parameter optimization. Schwartz and Nishino [ 23] introduced material traits that incorporate learned convolutional auto-encoder features. Recently, Cimpoi et al. [3] developed a CNN and improved Fisher vector (IFV) classifier that achieves state-of-the-art results on FMD and KTH-TIPS2. Finally, it has been shown that jointly predicting objects and materials can improve performance [10, 33]. Convolutional neural networks. While CNNs have been

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

8

Finally, it has been shown that jointly predicting objects and materials can improve performance [10, 33]. Convolutional neural networks. While CNNs have been around for a few decades, with early successes such as LeNet [ 14], they have only recently led to state-of-theart results in object classification and detection, leading to enormous progress. Driven by the ILSVRC challenge [ 21], we have seen many successful CNN architectures [ 32,24,28,27], led by the work of Krizhevsky et al. on their SuperVision (a.k.a. AlexNet ) network [ 13], with more recent architectures including GoogLeNet [ 28]. In addition to image classification, CNNs are the state-of-the-art for detection and localization of objects, with recent work including R-CNNs [ 7], Overfeat [ 24], and VGG [ 27]. Finally, relevant to our goal of per-pixel material segmentation, Farabet et al. [6] use a multi-scale CNN to predict the class at every pixel in a segmentation. Oquab et al. [18] employ a sliding window approach to localize patch classification of objects. We build on this body of work in deep learning to

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

9

sliding window approach to localize patch classification of objects. We build on this body of work in deep learning to solve our problem of material recognition and segmentation. 2 Brick Carpet Ceramic Fabric Foliage Food Glass Hair Leather Metal Mirror Other Painted Paper Plastic Pol. stone Skin Sky Stone Tile Wallpaper Water Wood Figure 2. Example patches from all 23 categories of the Materials in Context Database (MINC ). Note that we sample patches so that the patch center is the material in question (and not necessarily the entire patch). See Table 1 for the size of each category. 3. The Materials in Context Database (MINC) We now describe the methodology that went into building our new material database. Why a new database? We needed a dataset with the following properties: Size: It should be sufficiently large that learning methods can generalize beyond the training set. Well-sampled : Rare categories should be represented with a large number of examples. Diversity : Images should span a wide range of appearances of each material in real-world settings. Number of categories : It should contain many different materials found in the real world. 3.1. Sources of data We decided to start with the public, crowdsourced OpenSurfaces dataset [ 1] as the seed for MINC since it is

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

10

3.1. Sources of data We decided to start with the public, crowdsourced OpenSurfaces dataset [ 1] as the seed for MINC since it is drawn from Flickr imagery of everyday, real-world scenes with reasonable diversity. Furthermore, it has a large number of categories and the most samples of all prior databases. While OpenSurfaces data is a good start, it has a few limitations. Many categories in OpenSurfaces are not well sampled. While the largest category, wood , has nearly 20K samples, smaller categories, such as water , have only tens of examples. This imbalance is due to the way the OpenSurfaces dataset was annotated; workers on Amazon Mechanical Turk (AMT) were free to choose any material subregion to segment. Workers often gravitated towards certain common types of materials or salient objects, rather than being encouraged to label a diverse set of materials. Further, the images come from a single source (Flickr). We decided to augment OpenSurfaces with substantially more data, especially for underrepresented material cate-gories, with the initial goal of gathering at least 10K samples per material category. We decided to gather this data from another source of imagery, professional photos on the interior design website Houzz (houzz.com). Our motivation for using this different source of data was that, despite Houzz

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

11

another source of imagery, professional photos on the interior design website Houzz (houzz.com). Our motivation for using this different source of data was that, despite Houzz photos being more “staged” (relative to Flickr photos), they actually represent a larger variety of materials. For instance, Houzz photos contain a wide range of types of polished stone. With these sources of image data, we now describe how we gather material annotations. 3.2. Segments, Clicks, and Patches What specific kinds of material annotations make for a good database? How should we collect these annotations? The type of annotations to collect is guided in large part by the tasks we wish to generate training data for. For some tasks such as scene recognition, whole-image labels can suffice [ 31,34]. For object detection, labeled bounding boxes as in PASCAL are often used [ 5]. For segmentation or scene parsing tasks, per-pixel segmentations are required [ 22, 8]. Each style of annotation comes with a cost proportional to its complexity. For materials, we decided to focus on two problems, guided by prior work: Patch material classification. Given an image patch, what kind of material is it at the center?

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

12

problems, guided by prior work: Patch material classification. Given an image patch, what kind of material is it at the center? Full scene material classification. Given a full image, produce a full per-pixel segmentation and labeling. Also known as semantic segmentation orscene parsing (but in our case, focused on materials). Note that classification can be a component of segmentation, e.g., with sliding window approaches. 3 (a) Which images contain wood ?(b) Click on 3 points of wood(c) What is this material?Figure 3. AMT pipeline schematic for collecting clicks. (a) Workers filter by images that contain a certain material, (b)workers click on materials, and (c)workers validate click locations by re-labeling each point. Example responses are shown in orange. Segments. OpenSurfaces contains material segmentations— carefully drawn polygons that enclose same-material regions. To form the basis of MINC, we selected OpenSurfaces segments with high confidence (inter-worker agreement) and manually curated segments with low confidence, giving a total of 72K shapes. To better balance the categories, we

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

13

manually curated segments with low confidence, giving a total of 72K shapes. To better balance the categories, we manually segmented a few hundred extra samples for sky, foliage andwater . Since some of the OpenSurfaces categories are difficult for humans, we consolidated these categories. We found that many AMT workers could not disambiguate stone from concrete ,clear plastic from opaque plastic , and granite from marble . Therefore, we merged these into stone ,plastic , andpolished stone respectively. Without this merging, many ground truth examples in these categories would be incorrect. The final list of 23 categories is shown in Table 1. The category other is different in that it was created by combining various smaller categories. Clicks. Since we want to expand our dataset to millions of samples, we decided to augment OpenSurfaces segments by collecting clicks : single points in an image along with a material label, which are much cheaper and faster to collect. Figure 3 shows our pipeline for collecting clicks. Initially, we tried asking workers to click on examples of a given material in a photo. However, we found that workers would get frustrated if the material was absent in too many of the photos. Thus, we added an initial first

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

14

workers would get frustrated if the material was absent in too many of the photos. Thus, we added an initial first stage where workers filter out such photos. To increase the accuracy of our labels, we verify the click labels by asking different workers to specify the material for each click without providing them with the label from the previous stage. To ensure that we obtain high quality annotations and avoid collecting labels from workers who are not making an effort, we include secret known answers (sentinels) in the first and third stages, and block workers with an accuracy below 50% and 85% respectively. We do not use sentinels in the second stage since it would require per-pixel ground truth labels, and it turned out not to be necessary. Workers generally performed all three tasks so we could identify bad workers in the first or third task.Patches Category Patches Category Patches Category 564,891 Wood 114,085 Polished stone 35,246 Skin 465,076 Painted 98,891 Carpet 29,616 Stone 397,982 Fabric 83,644 Leather 28,108 Ceramic 216,368 Glass 75,084 Mirror 26,103 Hair 188,491 Metal 64,454 Brick 25,498 Food 147,346 Tile 55,364 Water 23,779 Paper

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

15

216,368 Glass 75,084 Mirror 26,103 Hair 188,491 Metal 64,454 Brick 25,498 Food 147,346 Tile 55,364 Water 23,779 Paper 142,150 Sky 39,612 Other 14,954 Wallpaper 120,957 Foliage 38,975 Plastic Table 1. MINC patch counts by category. Patches were created from both OpenSurfaces segments and our newly collected clicks . See Section 3.2 for details. Material clicks were collected for both OpenSurfaces images and the new Houzz images. This allowed us to use labels from OpenSurfaces to generate the sentinel data; we included 4 sentinels per task. With this streamlined pipeline we collected 2,341,473 annotations at an average cost of $0.00306 per annotation (stage 1: $0.02 / 40 images, stage 2: $0.10 / 50 images, 2, stage 3: $0.10 / 50 points). Patches. Labeled segments and clicks form the core of MINC . For training CNNs and other types of classifiers, it is useful to have data in the form of fixed-sized patches. We convert both forms of data into a unified dataset format: square image patches. We use a patch center andpatch scale

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

16

We convert both forms of data into a unified dataset format: square image patches. We use a patch center andpatch scale (a multiplier of the smaller image dimension) to define the image subregion that makes a patch. For our patch classification experiments, we use 23.3% of the smaller image dimension. Increasing the patch scale provides more context but reduces the spatial resolution. Later in Section 5 we justify our choice with experiments that vary the patch scale for AlexNet. We place a patch centered around each click label. For each segment , if we were to place a patch at every interior pixel then we would have a very large and redundant dataset. Therefore, we Poisson-disk subsample each segment , separating patch centers by at least 9.1% of the smaller image dimension. These segments generated 655,201 patches (an average of 9.05 patches per segment ). In total, we generated 2,996,674 labeled patches from 436,749 images. Patch counts are shown in Table 1, and example patches from various categories are illustrated in Figure 2. 4. Material recognition in real-world images Our goal is to train a system that recognizes the material at every pixel in an image. We split our training procedure into multiple stages and analyze the performance of the

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

17

Our goal is to train a system that recognizes the material at every pixel in an image. We split our training procedure into multiple stages and analyze the performance of the network at each stage. First, we train a CNN that produces a single prediction for a given input patch. Then, we convert the CNN into a sliding window and predict materials on a dense grid across the image. We do this at multiple scales and average to obtain a unary term. Finally, a dense CRF [12] combines the unary term with fully connected pairwise reasoning to output per-pixel material predictions. The entire system is depicted in Figure 1, and described more below. 4 Sky Water FoliageTile (a) Multiscale input (b) Probability map (12 of 23 shown) (c) Predicted labelsSliding CNNDense CRF Wood Plastic Sliding CNNSliding CNN Avg. Glass Hair Fabric Polished stone Stone PaperWaterSky Foliage TilePaperWood StoneWood TileFigure 4. Pipeline for full scene material classification. An image (a)is resized to multiple scales [1=p 2;1;p 2]. The same sliding CNN

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

18

2;1;p 2]. The same sliding CNN predicts a probability map (b)across the image for each scale; the results are upsampled and averaged. A fully connected CRF predicts a final label for each pixel (c). This example shows predictions from a single GoogLeNet converted into a sliding CNN (no average pooling). 4.1. Training procedure MINC contains 3 million patches that we split into training, validation and test sets. Randomly splitting would result in nearly identical patches (e.g., from the same OpenSurfaces segment) being put in training and test, thus inflating the test score. To prevent correlation, we group photos into clusters of near-duplicates, then assign each cluster to one of train, validate or test. We make sure that there are at least 75segments of each category in the test set to ensure there are enough segments to evaluate segmentation accuracy. To detect near-duplicates, we compare AlexNet CNN features computed from each photo (see the supplemental for details). For exact duplicates, we discard all but one of the copies. We train all of our CNNs by fine-tuning the network starting from the weights obtained by training on 1.2 million images from ImageNet (ILSVRC2012). When training

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

19

starting from the weights obtained by training on 1.2 million images from ImageNet (ILSVRC2012). When training AlexNet , we use stochastic gradient descent with batchsize 128, dropout rate 0.5, momentum 0.9, and a base learning rate of 103that decreases by a factor of 0.25 every 50,000 iterations. For GoogLeNet , we use batchsize 69, dropout 0.4, and learning rate t= 104p 1t=250000 for iteration t. Our training set has a different number of examples per class, so we cycle through the classes and randomly sample an example from each class. Failing to properly balance the examples results in a 5.7% drop in mean class accuracy (on the validation set). Further, since it has been shown to reduce overfitting, we randomly augment samples by taking crops (227227out of 256256), horizontal mirror flips, spatial scales in the range [1=p 2;p 2], aspect ratios from 3:4 to 4:3, and amplitude shifts in [0:95;1:05]. Since we are looking at local regions, we subtract a per-channel mean (R: 124, G:

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

20

and amplitude shifts in [0:95;1:05]. Since we are looking at local regions, we subtract a per-channel mean (R: 124, G: 117, B: 104) rather than a mean image [13]. 4.2. Full scene material classification Figure 4 shows an overview of our method for simultaneously segmenting and recognizing materials. Given a CNN that can classify individual points in the image, we convert it to a sliding window detector and densely classify a grid across the image. Specifically, we replace the last fully connected layers with convolutional layers, so that the network is fully convolutional and can classify images ofany shape. After conversion, the weights are fixed and not fine-tuned. With our converted network, the strides of each layer cause the network to output a prediction every 32 pixels. We obtain predictions every 16 pixels by shifting the input image by half-strides (16 pixels). While this appears to require 4x the computation, Sermanet et al. [24] showed that the convolutions can be reused and only the pool5 through fc8 layers need to be recomputed for the half-stride shifts. Adding half-strides resulted in a minor 0.2% improvement in mean class accuracy across segments (after applying the

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

21

Adding half-strides resulted in a minor 0.2% improvement in mean class accuracy across segments (after applying the dense CRF, described below), and about the same mean class accuracy at click locations. The input image is resized so that a patch maps to a 256x256 square. Thus, for a network trained at patch scale s, the resized input has smaller dimension d= 256=s. Note thatdis inversely proportional to scale, so increased context leads to lower spatial resolution. We then add padding so that the output probability map is aligned with the input when upsampled. We repeat this at 3 different scales (smaller dimensiond=p 2,d,dp 2), upsample each output probability map with bilinear interpolation, and average the predictions. To make the next step more efficient, we upsample the output to a fixed smaller dimension of 550. We then use the dense CRF of Kr¨ahenb ¨uhlet al. [12] to predict a label at every pixel, using the following energy: E(xjI) =X i i(xi) +X i<j ij(xi;xj) (1) i(xi) =logpi(xi) (2)

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

22

i i(xi) +X i<j ij(xi;xj) (1) i(xi) =logpi(xi) (2) ij(xi;xj) =wp(xi6=xj)k(fifj) (3) where iis the unary energy (negative log of the aggregated softmax probabilities) and ijis the pairwise term that connects every pair of pixels in the image. We use a single pairwise term with a Potts label compatibility term weighted by wpand unit Gaussian kernel k. For the featuresfi, we convert the RGB image to L*a*b*and use color (IL i;Ia i;Ib i)and position (px;py)as pairwise features for each pixel: fi=h px i pd;py i pd;IL i L;Ia i ab;Ib i abi , wheredis the smaller image dimension. Figure 4 shows an example unary termpiand the resulting segmentation x. 5 Figure 5. Varying patch scale . We train/test patches of different scales (the patch locations do not vary). The optimum is a trade-off

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

23

5 Figure 5. Varying patch scale . We train/test patches of different scales (the patch locations do not vary). The optimum is a trade-off between context and spatial resolution. CNN: AlexNet. Architecture Validation Test AlexNet [13] 82.2% 81.4% GoogLeNet [28] 85.9% 85.2% VGG-16 [27] 85.6% 84.8% Table 2. Patch material classification results. Mean class accuracy for different CNNs trained on MINC. See Section 5.1. Sky 97.3% Food 90.4% Wallpaper 83.4% Glass 78.5% Hair 95.3% Leather 88.2% Tile 82.7% Fabric 77.8% Foliage 95.1% Other 87.9% Ceramic 82.7% Metal 77.7% Skin 93.9% Pol. stone 85.8% Stone 82.7% Mirror 72.0% Water 93.6% Brick 85.1% Paper 81.8% Plastic 70.9% Carpet 91.6% Painted 84.2% Wood 81.3% Table 3. Patch test accuracy by category. CNN: GoogLeNet . See the supplemental material for a full confusion matrix. 5. Experiments and Results

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

24

Table 3. Patch test accuracy by category. CNN: GoogLeNet . See the supplemental material for a full confusion matrix. 5. Experiments and Results 5.1. Patch material classification In this section, we evaluate the effect of many different design decisions for training methods for material classification and segmentation, including various CNN architectures, patch sizes, and amounts of data. CNN Architectures. Our ultimate goal is full material segmentation, but we are also interested in exploring which CNN architectures give the best results for classifying single patches. Among the networks and parameter variations we tried we found the best performing networks were AlexNet [13],VGG-16 [27] and GoogLeNet [28].AlexNet andGoogLeNet are re-implementations by BVLC [ 11], and VGG-16 is configuration D (a 16 layer network) of [ 27]. All models were obtained from the Caffe Model Zoo [ 11]. Our experiments use AlexNet for evaluating material classification design decisions and combinations of AlexNet and GoogLeNet for evaluating material segmentation. Tables 2 and 3 summarize patch material classification results on our dataset. Figure 10 shows correct and incorrect predictions made with high confidence.

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

25

and 3 summarize patch material classification results on our dataset. Figure 10 shows correct and incorrect predictions made with high confidence. Input patch scale. To classify a point in an image we must decide how much context to include around it. The context, expressed as a fraction of image size, is the patch scale. A priori, it is not clear which scale is best since small patches have better spatial resolution, but large patches have more 0.5e6 1e6 1.5e6 2e6 2.5e6 Number of training patches0.600.650.700.750.800.85Mean class accuracy (test set)66.1%74.8%77.9%79.3%80.2%80.9% 81.2% 81.4%Equal size 79.6%Figure 6. Varying database size. Patch accuracy when trained on random subsets of MINC .Equal size is using equal samples per category (size determined by smallest category). CNN: AlexNet. Peak accuracy per category Figure 7. Accuracy vs patch scale by category . Dots: peak accuracy for each category; colored lines: sky,wallpaper ,mirror ; gray lines: other categories. CNN: AlexNet . While most materials are

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

26

lines: other categories. CNN: AlexNet . While most materials are optimally recognized at 23.3% or 32% patch scale, recognition of sky,wallpaper andmirror improve with increasing context. contextual information. Holding patch centers fixed we varied scale and evaluated classification accuracy with AlexNet . Results and a visualization of patch scales are shown in Figure 5. Scale 32% performs the best. Individual categories had peaks at middle scale with some exceptions; we find thatmirror ,wallpaper andskyimprove with increasing context (Figure 7). We used 23.3% (which has nearly the same accuracy but higher spatial resolution) for our experiments. Dataset size. To measure the effect of size on patch classification accuracy we trained AlexNet with patches from randomly sampled subsets of all 369,104 training images and tested on our full test set (Figure 6). As expected, using more data improved performance. In addition, we still have not saturated performance with 2.5 million training patches; even higher accuracies may be possible with more training data (though with diminishing returns). Dataset balance. Although we’ve shown that more data is better we also find that a balanced dataset is more effective. We trained AlexNet with all patches of our smallest category

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

27

better we also find that a balanced dataset is more effective. We trained AlexNet with all patches of our smallest category (wallpaper ) and randomly sampled the larger categories (the largest, wood , being 40x larger) to be equal size. We then measured mean class accuracy on the same full test set. As shown in Figure 6, “Equal size” is more accurate than a dataset of the same size and just 1.7% lower than the full training set (which is 9x larger). This result further demonstrates the value of building up datasets in a balanced manner, focusing on expanding the smallest, least common categories. 6 Brick WoodCarpet WaterWallpaperTileStoneSkySkinPolished stonePlasticPaperPainted OtherMirrorMetalLeatherHairGlassFoodFoliageFabricCeramicFigure 8. Full-scene material classification examples: high-accuracy test set predictions by our method. CNN: GoogLeNet (with the average pooling layer removed). Right: legend for material colors. See Table 4 for quantitative evaluation. Input image (a)Labels from CRF (b)Labels from CRF (test set) trained on segments trained on clicks

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

28

Input image (a)Labels from CRF (b)Labels from CRF (test set) trained on segments trained on clicks Figure 9. Optimizing for click accuracy leads to sloppy boundaries. In(a), we optimize for mean class accuracy across segments, resulting in high quality boundaries. In (b), we optimize for mean class accuracy at click locations. Since the clicks are not necessarily close to object boundaries, there is no penalty for sloppy boundaries. CNN: GoogLeNet (without average pooling). 5.2. Full scene material segmentation The full test set for our patch dataset contains 41,801 photos, but most of them contain only a few labels. Since we want to evaluate the per-pixel classification performance, we select a subset of 5,000 test photos such that each photo contains a large number of segments and clicks, and small categories are well sampled. We greedily solve for the best such set of photos. We similarly select 2,500 of 25,844 validation photos. Our splits for all experiments are included online with the dataset. To train the CRF for our model, we try various parameter settings ( p,ab,L,wp) and select the model that performs best on the validation set. In total, we

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

29

try various parameter settings ( p,ab,L,wp) and select the model that performs best on the validation set. In total, we evaluate 1799 combinations of CNNs and CRF parameters. See the supplemental material for a detailed breakdown. We evaluate multiple versions of GoogLeNet : both the original architecture and a version with the average pooling layer (at the end) changed to 5x5, 3x3, and 1x1 (i.e. no average pooling). We evaluate AlexNet trained at multiple patch scales (Figure 5). When using an AlexNet trained at a different scale, we keep the same scale for testing. We also experiment with ensembles of GoogLeNet andAlexNet ,Architecture(a)Segments only (b)Clicks only Class Total Class Total AlexNet Scale: 11.6% 64.3% 72.6% 79.9% 77.2% AlexNet Scale: 23.3% 69.6% 76.6% 83.3% 81.1% AlexNet Scale: 32.0% 70.1% 77.1% 83.2% 80.7% AlexNet Scale: 46.5% 69.6% 75.4% 80.8% 77.7%

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

30

AlexNet Scale: 46.5% 69.6% 75.4% 80.8% 77.7% AlexNet Scale: 66.2% 67.7% 72.0% 77.2% 72.6% GoogLeNet 7x7 avg. pool 64.4% 71.6% 63.6% 63.4% GoogLeNet 5x5 avg. pool 67.6% 74.6% 70.9% 69.8% GoogLeNet 3x3 avg. pool 70.4% 77.7% 76.1% 74.7% GoogLeNet No avg. pool 70.4% 78.8% 79.1% 77.4% Ensemble 2 CNNs 73.1% 79.8% 84.5% 83.1% Ensemble 3 CNNs 73.1% 79.3% 85.9% 83.5% Ensemble 4 CNNs 72.1% 78.4% 85.8% 83.2% Ensemble 5 CNNs 71.7% 78.3% 85.5% 83.2% Table 4. Full scene material classification results. Mean class and total accuracy on the test set. When training, we optimize the CRF parameters for mean class accuracy, but report both mean class

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

31

and total accuracy on the test set. When training, we optimize the CRF parameters for mean class accuracy, but report both mean class and total accuracy (mean accuracy across all examples). In one experiment (a), we train and test only on segments; in a separate experiment (b), we train and test only on clicks. Accuracies for segments are averaged across all pixels that fall in that segment. combined with either arithmetic or geometric mean. Since we have two types of data, clicks andsegments , we run two sets of experiments: (a) we train and test only on segments, and in a separate experiment (b) we train and test only on clicks. These two training objectives result in very different behavior, as illustrated in Figure 9. In experiment (a), the accuracy across segments are optimized, producing clean boundaries. In experiment (b), the CRF maximizes accuracy only at click locations, thus resulting in sloppy boundaries. As shown in Table 4, the numerical scores for the two experiments are also very different: segments are more challenging than clicks. While clicks are sufficient to train a CNN, they are not sufficient to train a CRF. Focusing on segmentation accuracy, we see from Table 4(a) that our best single model is GoogLeNet without

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

32

Focusing on segmentation accuracy, we see from Table 4(a) that our best single model is GoogLeNet without average pooling (6% better than with pooling). The best ensemble is 2 CNNs: GoogLeNet (no average pooling) and AlexNet (patch scale: 46.5%), combined with arithmetic mean. Larger ensembles perform worse since we are aver7 Correct Fabric (99%) Foliage (99%) Food (99%) Leather (99%)Correct Metal (99%) Mirror (99%) Painted (99%) Plastic (99%)Incorrect T: Wood T: Polished stone T: Water T: Fabric P: Stone (90%) P: Water (35%) P: Carpet (27%) P: Foliage (67%) Figure 10. High confidence predictions. Top two rows: correct predictions. Bottom row: incorrect predictions ( T: true, P: predicted). Percentages indicate confidence (the predictions shown are at least this confident). CNN: GoogLeNet. aging worse CNNs. In Figure 8, we show example labeling results on test images. 5.3. Comparing MINC to FMD Compared to FMD, the size and diversity of MINC is

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

33

results on test images. 5.3. Comparing MINC to FMD Compared to FMD, the size and diversity of MINC is valuable for classifying real-world imagery. Table 5 shows the effect of training on all of FMD and testing on MINC (and vice versa). The results suggests that training on FMD alone is not sufficient for real-world classification. Though it may seem that our dataset is “easy,” since the best classifications scores are lower for FMD than for MINC , we find that difficulty is in fact closely tied to dataset size (Section 5.1). Taking 100 random samples per category, AlexNet achieves 54.20.7% on MINC (64.6 1.3% when considering only the 10 FMD categories) and 66.5% on FMD. 5.4. Comparing CNNs with prior methods Cimpoi [ 3] is the best prior material classification method on FMD. We find that by replacing DeCAF with oversampled AlexNet features we can improve on their FMD results. We then show that on MINC , a finetuned CNN is even better.

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

34

We then show that on MINC , a finetuned CNN is even better. To improve on [ 3], we take their SIFT IFV, combine it with AlexNet fc7 features, and add oversampling [ 13] (see supplemental for details). With a linear SVM we achieve 69.60.3% on FMD. Previous results are listed in Table 6. Having found that SIFT IFV+fc7 is the new best on FMD, we compare it to a finetuned CNN on a subset of MINC (2500 patches per category, one patch per photo). Fine-tuning AlexNet achieves 76.00.2% whereasTest FMD MINC TrainFMD 66.5% 26.1% MINC 41.7% 85.0%(10 categories in common) Table 5. Cross-dataset experiments. We train on one dataset and test on another dataset. Since MINC contains 23 categories, we limit MINC to the 10 categories in common. CNN: AlexNet. Method Accuracy Trials Sharan et al. [25] 57.10.6% 14 splits Cimpoi et al. [3] 67.10.4% 14 splits Fine-tuned AlexNet 66.51.5% 5 folds

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

35

Cimpoi et al. [3] 67.10.4% 14 splits Fine-tuned AlexNet 66.51.5% 5 folds SIFT IFV+fc7 69.60.3% 10 splits Table 6. FMD experiments. By replacing DeCAF features with oversampled AlexNet features we improve on the best FMD result. SIFT IFV+fc7 achieves 67.4 0.5% with a linear SVM (oversampling, 5 splits). This experiment shows that a finetuned CNN is a better method for MINC than SIFT IFV+fc7. 6. Conclusion Material recognition is a long-standing, challenging problem. We introduce a new large, open, material database, MINC , that includes a diverse range of materials of everyday scenes and staged designed interiors, and is at least an order of magnitude larger than prior databases. Using this large database we conduct an evaluation of recent deep learning algorithms for simultaneous material classification and segmentation, and achieve results that surpass prior attempts at material recognition. Some lessons we have learned are: Training on a dataset which includes the surrounding context is crucial for real-world material classification.

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

36

at material recognition. Some lessons we have learned are: Training on a dataset which includes the surrounding context is crucial for real-world material classification. Labeled clicks are cheap and sufficient to train a CNN alone. However, to obtain high quality segmentation results, training a CRF on polygons results in much better boundaries than training on clicks. Many future avenues of work remain. Expanding the dataset to a broader range of categories will require new ways to mine images that have more variety, and new annotation tasks that are cost-effective. Inspired by attributes for textures [ 3], in the future we would like to identify material attributes and expand our database to include them. We also believe that further exploration of joint material and object classification and segmentation will be fruitful [ 10] and lead to improvements in both tasks. Our database, trained models, and all experimental results are available online at http://minc.cs.cornell.edu/. Acknowledgements. This work was supported in part by Google, Amazon AWS for Education, a NSERC PGS-D scholarship, the National Science Foundation (grants IIS1149393, IIS-1011919, IIS-1161645), and the Intel Science and Technology Center for Visual Computing. 8 References

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

37

and Technology Center for Visual Computing. 8 References [1]S. Bell, P. Upchurch, N. Snavely, and K. Bala. OpenSurfaces: A richly annotated catalog of surface appearance. ACM Trans. on Graphics (SIGGRAPH) , 32(4), 2013. 1, 2, 3 [2]B. Caputo, E. Hayman, and P. Mallikarjuna. Class-specific material categorisation. In ICCV , pages 1597–1604, 2005. 2 [3]M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, and A. Vedaldi. Describing textures in the wild. In CVPR , pages 3606–3613. IEEE, 2014. 2, 8 [4]K. J. Dana, B. Van Ginneken, S. K. Nayar, and J. J. Koenderink. Reflectance and texture of real-world surfaces. ACM Transactions on Graphics (TOG) , 18(1):1–34, 1999. 2 [5]M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The Pascal Visual Object Classes (VOC)

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

38

and A. Zisserman. The Pascal Visual Object Classes (VOC) Challenge. IJCV , 88(2):303–338, June 2010. 3 [6]C. Farabet, C. Couprie, L. Najman, and Y . LeCun. Learning hierarchical features for scene labeling. PAMI , 35(8):1915– 1929, 2013. 2 [7]R. Girshick, J. Donahue, T. Darrell, and J. Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In CVPR , 2014. 2 [8]S. Gould, R. Fulton, and D. Koller. Decomposing a scene into geometric and semantically consistent regions. In ICCV , 2009. 3 [9]E. Hayman, B. Caputo, M. Fritz, and J. olof Eklundh. On the significance of real-world conditions for material classification. In ECCV , 2004. 2 [10] D. Hu, L. Bo, and X. Ren. Toward robust material recognition for everyday objects. In BMVC , pages 1–11. Citeseer, 2011. 1, 2, 8

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

39

for everyday objects. In BMVC , pages 1–11. Citeseer, 2011. 1, 2, 8 [11] Y . Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S. Guadarrama, and T. Darrell. Caffe: Convolutional architecture for fast feature embedding. In Proceedings of the ACM International Conference on Multimedia , pages 675– 678. ACM, 2014. 6 [12] P. Kr ¨ahenb ¨uhl and V . Koltun. Parameter learning and convergent inference for dense random fields. In ICML , pages 513–521, 2013. 2, 4, 5 [13] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems , pages 1097–1105, 2012. 2, 5, 6, 8 [14] Y . LeCun, B. Boser, J. S. Denker, D. Henderson, R. E. Howard, W. Hubbard, and L. D. Jackel. Backpropagation

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

40

Howard, W. Hubbard, and L. D. Jackel. Backpropagation applied to handwritten zip code recognition. Neural computation, 1(4):541–551, 1989. 2 [15] T. Leung and J. Malik. Representing and recognizing the visual appearance of materials using three-dimensional textons. IJCV , 43(1):29–44, June 2001. 2 [16] W. Li and M. Fritz. Recognizing materials from virtual examples. In ECCV , pages 345–358. Springer, 2012. 2 [17] C. Liu, L. Sharan, E. H. Adelson, and R. Rosenholtz. Exploring features in a bayesian framework for material recognition. InCVPR , pages 239–246. IEEE, 2010. 1, 2 [18] M. Oquab, L. Bottou, I. Laptev, and J. Sivic. Learning and transferring mid-level image representations using convolutional neural networks. In CVPR , 2014. 2[19] G. Patterson, C. Xu, H. Su, and J. Hays. The SUN Attribute Database: Beyond Categories for Deeper Scene Understanding. IJCV , 108(1-2):59–81, 2014. 1

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

41

Database: Beyond Categories for Deeper Scene Understanding. IJCV , 108(1-2):59–81, 2014. 1 [20] X. Qi, R. Xiao, J. Guo, and L. Zhang. Pairwise rotation invariant co-occurrence local binary pattern. In ECCV , pages 158–171. Springer, 2012. 1, 2 [21] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, A. C. Berg, and L. Fei-Fei. ImageNet Large Scale Visual Recognition Challenge, 2014. 1, 2 [22] B. C. Russell, A. Torralba, K. P. Murphy, and W. T. Freeman. LabelMe: A database and web-based tool for image annotation. IJCV , 77(1-3):157–173, May 2008. 3 [23] G. Schwartz and K. Nishino. Visual material traits: Recognizing per-pixel material context. In Proceedings of the International Conference on Computer Vision Workshops (ICCVW) , pages 883–890. IEEE, 2013. 2

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

42

International Conference on Computer Vision Workshops (ICCVW) , pages 883–890. IEEE, 2013. 2 [24] P. Sermanet, D. Eigen, X. Zhang, M. Mathieu, R. Fergus, and Y . LeCun. Overfeat: Integrated recognition, localization and detection using convolutional networks. In International Conference on Learning Representations (ICLR2014) . CBLS, April 2014. 2, 5 [25] L. Sharan, C. Liu, R. Rosenholtz, and E. Adelson. Recognizing materials using perceptually inspired features. IJCV , 2013. 1, 8 [26] L. Sharan, R. Rosenholtz, and E. Adelson. Material perception: What can you see in a brief glance? Journal of Vision , 9(8):784–784, 2009. 1, 2 [27] K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 , 2014. 2, 6 [28] C. Szegedy, W. Liu, Y . Jia, P. Sermanet, S. Reed, D. Anguelov,

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

43

[28] C. Szegedy, W. Liu, Y . Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V . Vanhoucke, and A. Rabinovich. Going deeper with convolutions. CVPR , 2015. 2, 6 [29] R. Timofte and L. J. Van Gool. A training-free classification framework for textures, writers, and materials. In BMVC , pages 1–12, 2012. 2 [30] M. Varma and A. Zisserman. A statistical approach to texture classification from single images. IJCV , 62(1-2):61–81, Apr. 2005. 2 [31] J. Xiao, K. A. Ehinger, J. Hays, A. Torralba, and A. Oliva. SUN Database: Exploring a large collection of scene categories. IJCV , 2014. 1, 3 [32] M. D. Zeiler and R. Fergus. Visualizing and understanding convolutional networks. In ECCV , pages 818–833. Springer, 2014. 2 [33] S. Zheng, M.-M. Cheng, J. Warrell, P. Sturgess, V . Vineet,

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.0623

44

2014. 2 [33] S. Zheng, M.-M. Cheng, J. Warrell, P. Sturgess, V . Vineet, C. Rother, and P. H. Torr. Dense semantic image segmentation with objects and attributes. In CVPR , pages 3214–3221. IEEE, 2014. 2 [34] B. Zhou, A. Lapedriza, J. Xiao, A. Torralba, and A. Oliva. Learning deep features for scene recognition using Places database. NIPS , 2014. 1, 3 9

1412.0623

Material Recognition in the Wild with the Materials in Context Database

Recognizing materials in real-world images is a challenging task. Real-world materials have rich surface texture, geometry, lighting conditions, and clutter, which combine to make the problem particularly difficult. In this paper, we introduce a new, large-scale, open dataset of materials in the wild, the Materials in Context Database (MINC), and combine this dataset with deep learning to achieve material recognition and segmentation of images in the wild. MINC is an order of magnitude larger than previous material databases, while being more diverse and well-sampled across its 23 categories. Using MINC, we train convolutional neural networks (CNNs) for two tasks: classifying materials from patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC we found that the best performing CNN architectures can achieve 85.2% mean class accuracy. We convert these trained CNN classifiers into an efficient fully convolutional framework combined with a fully connected conditional random field (CRF) to predict the material at every pixel in an image, achieving 73.1% mean class accuracy. Our experiments demonstrate that having a large, well-sampled dataset such as MINC is crucial for real-world material recognition and segmentation.

http://arxiv.org/pdf/1412.0623

CVPR 2015. Sean Bell and Paul Upchurch contributed equally

cs.CV

20141201

20150414

1412.1123

0

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection Gedas Bertasius University of Pennsylvania gberta@seas.upenn.eduJianbo Shi University of Pennsylvania jshi@seas.upenn.eduLorenzo Torresani Dartmouth College lt@dartmouth.edu Abstract Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit objectrelated features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.1123

1

lifted from a pre-trained network optimized over a largescale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independentlytrained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multiscale deep learning approach achieves state-of-the-art results in contour detection. 1. Introduction Contour detection is typically considered a low-level problem, and used to aid higher-level tasks such as object detection [1, 3, 31, 24]. However, it can be argued that the tasks of detecting objects and predicting contours are closely related. For instance, given the contours we can easily infer which objects are present in the image. Con0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 100.10.20.30.40.50.60.70.80.91 RecallPrecision [F=.80] Human

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.1123

2

RecallPrecision [F=.80] Human [F=.753] DeepEdge [F=.753] N4−fields [10] [F=.747] MCG [2] [F=.746] SE [8] [F=.739] SCG [27] [F=.737] PMI [14] [F=.727] Sketch Tokens [19] [F=.726] gPb−owt−ucm [1]Figure 1: Contour detection accuracy on the BSDS500 dataset. Our method attains higher average precision compared to prior methods and state-of-the-art F-score. At low recall, DeepEdge achieves nearly 100% precision.. versely, if we are given exact locations of the objects we could predict contours just as easily. A commonly established pipeline in computer vision starts with with low-level contour prediction and then moves up to higher-level object detection. However, since we claim that these two tasks are mutually related, we propose to invert this process. Instead of using contours as low-level cues for object detection, we want to use object-specific information as high-level cues for contour detection. Thus, in a sense our scheme can be

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.1123

3

want to use object-specific information as high-level cues for contour detection. Thus, in a sense our scheme can be viewed as a top-down approach where object-level cues inform the low-level contour detection process. In this work, we present a unified multi-scale deep learning approach that uses higher-level object information to 1arXiv:1412.1123v3 [cs.CV] 23 Apr 2015 predict contours. Specifically, we present a front-to-end convolutional architecture where contours are learned directly from raw pixels. Our proposed deep learning architecture reuses features computed by the first five convolutional layers of the network of Krizhevsky et al. [17]. We refer to this network as the KNet . Because the KNet has been trained for object classification, reusing its features enable our method to incorporate object-level information for contour prediction. In the experimental section, we will show that this high-level object information greatly improves contour detection results. Furthermore, our defined architecture operates on multiple scales simultaneously and combines local and global information from the image, which leads to significantly improved contour detection accuracy rates.

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.1123

4

information from the image, which leads to significantly improved contour detection accuracy rates. We connect the features computed by the convolutional layers of the KNet at four different scales of the input with alearned subnetwork that bifurcates into two branches (the architecture of our model is illustrated in Fig. 2). What should the learning objective be? When a human observer decides if a pixel is a boundary edge, a number of supporting evidence is used with object level reasoning. While it is impossible to record such information, we do have the fraction of observers in agreement for each pixel. We argue that a learning objective that predicts the fraction of human labelers in agreement can mimic human reasoning better. Thus, in the bifurcated sub-network we optimize the two branches with different learning objectives. The weights in one branch are optimized with an edge classification objective, while the other branch is trained to predict the fraction of human labelers in agreement, i.e., using a regression criterion. We show that predictions from the classification branch yield high edge recall, while the outputs of the regression branch have high precision. Thus, fusing these two outputs allows us to obtain excellent results with respect to

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.1123

5

branch yield high edge recall, while the outputs of the regression branch have high precision. Thus, fusing these two outputs allows us to obtain excellent results with respect to both metrics and produce state-of-the-art F-score and average precision. In summary, the use of higher-level object features, independent optimization of edge classification and regression objectives, as well as a unified multi-scale architecture are the key characteristics that allow our method to achieve the state-of-the-art in contour detection (see Fig. 1). 2. Related Work Deep Learning. In the recent years, deep convolutional networks have achieved remarkable results in a wide array of computer vision tasks [32, 25, 33, 17]. However, thus far, applications of convolutional networks focused on highlevel vision tasks such as face recognition, image classification, pose estimation or scene labeling [32, 25, 33, 17]. Excellent results in these tasks beg the question whether convolutional networks could perform equally well in lowerFigure 2: Visualization of multi-scale DeepEdge network architecture. To extract candidate contour points, we run the Canny edge detector. Then, around each candidate point, we extract patches at four different scales and simultaneously run them through the five convolutional layers of the

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.1123

6

Canny edge detector. Then, around each candidate point, we extract patches at four different scales and simultaneously run them through the five convolutional layers of the KNet [17]. We connect these convolutional layers to two separately-trained network branches. The first branch is trained for classification, while the second branch is trained as a regressor. At testing time, the scalar outputs from these two sub-networks are averaged to produce the final score. level vision tasks such as contour detection. In this paper, we present a convolutional architecture that achieves stateof-the-art results in a contour detection task, thus demonstrating that convolutional networks can be applied successfully for lower-level vision tasks as well. Edge Detection. Most of the contour detection methods can be divided into two branches: local and global methods. Local methods perform contour detection by reasoning about small patches inside the image. Some recent local methods include sketch tokens [18] and structured edges [7], Both of these methods are trained in a supervised fashion using a random forest classifier. Sketch tokens [18] pose contour detection as a multi-class classification task and predicts a label for each of the pixels individually. Structured edges [7], on the other hand, attempt to

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.1123

7

predict the labels of multiple pixels simultaneously. Global methods predict contours based on the information from the full image. Some of the most successful approaches in this genre are the MCG detector [2], gPb detector [1] and sparse code gradients [26]. While sparse code gradients use supervised SVM learning [4], both gPb and MCG rely on some form of spectral methods. Other spectral-based methods include Normalized Cuts [30] and PMI [13]. Recently, there have also been attempts to apply deep learning methods to the task of contour detection. While SCT [20] is a sparse coding approach, both N4fields [9] and DeepNet [16] use Convolutional Neural Networks (CNNs) to predict contours. N4fields rely on dictionary a) Input Image c)2ndlayer d)3rdlayer e)4thlayer e)5thlayer Figure 3: Visualization of the activation values at the selected convolutional filters of the KNet (filters are resized to the original image dimensions). The filters in the second layer fire on oriented edges inside the image. The third and fourth

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.1123

8

original image dimensions). The filters in the second layer fire on oriented edges inside the image. The third and fourth convolutional layers produce an outline of the shape of the object. The fifth layer fires on the specific parts of the object. learning and the use of the Nearest Neighbor algorithm within a CNN framework while DeepNet uses a traditional CNN architecture to predict contours. In comparison to these prior approaches, our work offers several contributions. First, we define a novel multi-scale bifurcated CNN architecture that enables our network to achieve state-of-the-art contour detection results. Second, we avoid manual feature engineering by learning contours directly from raw data. Finally, we believe that we are the first to propose the use of high-level object features for contour detection, thus inverting the traditional pipeline relying on low-level cues for mid-level feature extraction. Our experiments show that this top-down approach for contour detection yields state-of-the-art results. 3. The DeepEdge Network In this section, we describe our proposed deep learning approach for contour detection. For simplicity, we first

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.1123

9

3. The DeepEdge Network In this section, we describe our proposed deep learning approach for contour detection. For simplicity, we first present our architecture in the single-scale scenario (subsection 3.1) and then discuss how to take advantage of multiple scales (subsection 3.2). 3.1. Single-Scale Architecture Selection of Candidate Edge Points. To extract a set of candidate contours with high recall we apply the Canny edge detector [5] to the input image. For each of these points we then extract a patch of fixed size such that our candidate point is the center of the patch. Patches that do not fit into the image boundaries are padded with the mirror reflections of itself. Extraction of High-Level Features. We then resize the patch of fixed size to match the input dimensions of the KNet [17] and use this network to extract object-level features. The KNet is an appropriate model for our setting as it has been trained over a large number of object classes (the 1000 categories of the ImageNet dataset [28]) and thus captures features that are generic and useful for many object categories. While such features have been optimized for the task of object class recognition, they have been shown to

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.1123

10

categories. While such features have been optimized for the task of object class recognition, they have been shown to be highly effective for other image-analysis tasks, includ-ing object detection [10], attribute prediction [34], and image style recognition [15]. The network was trained on 1:2 million images and it includes more than 60million parameters. Its architecture consists of 5convolutional layers and 3fully connected layers. As we intend to use the KNet as a feature extractor for boundary detection, we utilize only the first 5convolutional layers, which preserve explicit location information before the “spatial scrambling” of the fully connection layers (note that a spatially variant representation is crucially necessary to predict the presence of contours at individual pixels). The first two KNet convolutional layers learn low-level information. As we move into the deeper layers, however, we observe that the network learns higher-level object information. The second convolutional layer seems to encode coherent edge structures. The third convolutional layer fires at locations corresponding to prototypical object shapes. The fourth layer appears to generate high responses for full shapes of the object, whereas the fifth layer fires on the specific object parts (See Fig. 3).

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.1123

11

the fifth layer fires on the specific object parts (See Fig. 3). In order to obtain a representation that captures this hierarchical information, we perform feature extraction at each of the five convolutional layers, as shown in Fig. 5. Specifically, we consider a small sub-volume of the feature map stack produced at each layer. The sub-volume is centered at the center of the patch in order to assess the presence of a contour in a small area around the candidate point. We then perform max,average , and center pooling on this subvolume. This yields a feature descriptor of size 3Fwhere Fis the number of feature maps computed by the convolutional layer. While max and average pooling are well established operations in deep learning, we define center pooling as selecting the center-value from each of the feature maps. The motivation for center pooling is that for each candidate point we want to predict the contour presence at that particular point. Because the candidate point is located at the center of the input patch, center pooling extracts the activation value from the location that corresponds to our candidate point location. A Bifurcated Sub-Network. We connect the feature maps computed via pooling from the five convolutional

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.1123

12

point location. A Bifurcated Sub-Network. We connect the feature maps computed via pooling from the five convolutional layers to two separately-trained network branches. Each Figure 4: A few samples of ground truth data illustrating the difference between the classification (first row) and the regression (second row) objectives. The classification branch is trained to detect contours that are marked by at least one of the human annotators. Conversely, the regression branch is optimized to the contour values that depict the fraction of human annotators agreeing on the contour. branch consists of two fully-connected layers. The first branch is trained using binary labels, i.e., to perform contour classification. This branch is making less selective predictions by classifying whether a given point is a contour or not. In a sense, this classification branch abstracts details related to the edge structure (orientation, strength, etc) and simply tries to predict the presence/absence of an edge at a given point. Due to such abstractions, the classification branch produces contour predictions with high recall. The second branch is optimized as a regressor to predict

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.1123

13

a given point. Due to such abstractions, the classification branch produces contour predictions with high recall. The second branch is optimized as a regressor to predict the fraction of human labelers agreeing about the contour presence at a particular point. Due to a regression objective, this branch is much more selective than the first branch. Intuitively, the second branch is trained to learn the structural differences between the contours that are marked by a different fraction of human labelers. For instance, the area that was labeled as a contour by 80% of human labelers must be significantly different than the area that was labeled as a contour by 20% human labelers. The regression branch is trying to learn such differences by predicting the fraction of human labelers who would mark a particular point as a contour. Thus, in a sense, we are training the regression branch to implicitly mimic how human labelers reason about the contour presence at a given point. In the experimental section, we demonstrate that due to its selectivity, the regression branch produces contour predictions with very high precision. In Fig. 4, we present several samples of ground truth data that illustrate the different properties of our two end-objectives. The number of hidden layers in the first and second fully

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.1123

14

of ground truth data that illustrate the different properties of our two end-objectives. The number of hidden layers in the first and second fully connected layers of both branches are 1024 and512, respectively. Both branches optimize the sum of squared difference loss over the (binary or continuous) labels. At test-ing time, the scalar outputs computed from these two subnetworks are averaged to produce a final score indicative of the probability that the candidate point is a contour. Visualization of this architecture is presented in Fig. 5. In order to train our sub-network, we generate patch examples and labels using training images with ground truth annotations from multiple human labelers. To generate the binary labels, we first sample 40;000positive examples that were marked as contours by at least one of the labelers. To generate negative examples we consider the points that were selected as candidate contour points by the Canny edge detector but that have not been marked as contours by any of the human labelers. These are essentially false positives. For training, we use a random subset of 40;000 of such points in order to have equal proportion of negative and positive examples. These 80;000examples are then used to train our classification sub-network.

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.1123

15

points in order to have equal proportion of negative and positive examples. These 80;000examples are then used to train our classification sub-network. In addition to the binary labels, we also generate continuous labels that are used to train the regression network. For this purpose, we define the regression label of a point to be the fraction of human labelers that marked the point as a contour. These 80;000examples with continuous labels are then also used to train our regression sub-network. 3.2. Multi-Scale Architecture In the previous section we presented a convolutional architecture for contour prediction utilizing a single scale. However, in practice, we found that a multi-scale approach works much better. In this section, we show how to modify the single-scale architecture so that it can exploit multiple scales simultaneously. Rather than extracting a patch at a single scale as we did in the previous section, in a multi-scale setting we extract |{z} |{z}Fixed WeightsLearned WeightsClassification Branch Regression Branch } }Figure 5: Detailed illustration of our proposed architecture in a single-scale setting. First, an input patch, centered around the

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.1123

16

Regression Branch } }Figure 5: Detailed illustration of our proposed architecture in a single-scale setting. First, an input patch, centered around the candidate point, goes through five convolutional layers of the KNet . To extract high-level features, at each convolutional layer we extract a small sub-volume of the feature map around the center point, and perform max,average , and center pooling on this sub-volume. The pooled values feed a bifurcated sub-network. At testing time, the scalar outputs computed from the branches of a bifurcated sub-networks are averaged to produce a final contour prediction. patches around the candidate point for different patch sizes so that they cover different spatial extents of the image. We then resize the patches to fit the KNet input and pump them in parallel through the five convolutional layers. Our highlevel features are then built by performing max, average and center pooling in a small sub-volume of the feature map at each convolutional layer and at each scale. This effectively increases the dimensionality of the feature vector by a factor equal to the number of scales compared to the single-scale setting. These pooled features are then connected as before

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.1123

17

increases the dimensionality of the feature vector by a factor equal to the number of scales compared to the single-scale setting. These pooled features are then connected as before to the two separately-trained network branches. A visualization of our multi-scale architecture is shown in Fig. 2. 3.3. Implementation Details In this section, we describe additional implementation details of our model. Our deep network is implemented using the software library Caffe [14]. We use four different scales for our patches. The sizes of these patches are 6464;128128;196196and a full-sized image. All of the patches are then resized to the KNet input dimensions of 227227. When extracting high-level features from the convolutional layers of KNet , we use sub-volumes of convolutionalfeature maps having spatial sizes 77;55;33;33, and33for the convolutional layers 1;2;3;4;5, respectively. Note that we shrink the size of the subvolume as we go deeper in the network since the feature maps get smaller due to pooling. Our choice of subvolume sizes is made to ensure we are roughly considering the same spatial extent of the original image at each layer. As illustrated in Fig. 5, during the training the weights

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.1123

18

ensure we are roughly considering the same spatial extent of the original image at each layer. As illustrated in Fig. 5, during the training the weights in the convolutional layers are fixed and only the weights in the fully connected layers of the two branches are learned. To train our model we use the learning rate of 0:1, the dropout fraction of 0:5,50number of epochs, and the size of the batch equal to 100. As described earlier, to train classification and regression branches we sample 80;000examples with binary labels. We also generate continuous labels for these 80;000 examples. In addition, we sample a hold-out dataset of size 40;000. This hold-out dataset is used for the hard-positive mining step [22]. For the first 25epochs we train classification and regression branches independently on the original 80;000sample training dataset. After the first 25epochs, we test both branches on the hold-out dataset and detect false negative predictions made by each branch. We then use these false negative examples along with the same number of randomly selected true negative examples to augment our original 80;000training dataset. For the remaining 25epochs, we train both branches on this augmented dataset.

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.1123

19

selected true negative examples to augment our original 80;000training dataset. For the remaining 25epochs, we train both branches on this augmented dataset. The motivation for the hard-positive mining step is to reduce the number of false negative predictions produced by both branches. By augmenting the original 80;000sized training data with false negative examples, we are forcing both branches to focus on hard positive examples, and thus, effectively reducing the number of false negative predictions. 4. Experiments In this section, we present our results on the BSDS500 dataset [23], which is arguably the most established benchmark for contour detection. This dataset contains 200training images, 100 validation images, and 200 test images. Contour detection accuracy is evaluated using three standard measures: fixed contour threshold (ODS), per-image best threshold (OIS), and average precision (AP). In section 4.1 we quantitatively compare our approach to the state-of-the-art. In sections 4.2-4.5 we study how the performance of our system changes as we modify some of the architecture choices (number of scales, feature maps, pooling scheme, training objective). This will cast additional insight into the factors that critically contribute to the high accuracy of our system. 4.1. Comparison with Other Methods

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.1123

20

pooling scheme, training objective). This will cast additional insight into the factors that critically contribute to the high accuracy of our system. 4.1. Comparison with Other Methods We compare the results produced by our approach and previously proposed contour detection methods. Table 1 summarizes the results. We note that our algorithm achieves contour detection accuracy that is higher or equal to stateof-the-art results according to two of the three metrics. Fig. 1 shows the precision and recall curve for the methods considered in our comparison. It also lists the F-score for each method (in the legend). We observe that there is the accuracy margin separating our approach from prior techniques. In particular, for low-recall our method achieves almost perfect precision rate. It also produces state-of-the-art F-score. 4.2. Single Scale versus Multiple Scales In this section we study the benefits of a multi-scale architecture. Results in Table 2 report accuracy for different numbers and choices of scales. The first four rows in the table illustrate the results achieved using a single-scale approach. Specifically, these four cases show performance obtained when training and testing our system with an input

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.1123

21

the table illustrate the results achieved using a single-scale approach. Specifically, these four cases show performance obtained when training and testing our system with an input patch of size 6464;128128;196196or a full-sized image, respectively. Note that adding information from multiple scales leads to significantly higher F-scores and higherMethod ODS OIS AP Felz., Hutt. [8] 0.610 0.640 0.560 Mean Shift [6] 0.640 0.680 0.560 Ncuts [30] 0.640 0.680 0.450 SCT [20] 0.710 0.720 0.740 gPb-owt-ucm [1] 0.726 0.757 0.696 Sketch Tokens [18] 0.727 0.746 0.780 PMI [13] 0.737 0.771 0.783 DeepNet [16] 0.738 0.759 0.758 SCG [26] 0.739 0.758 0.773 SE [7] 0.746 0.767 0.803 MCG [2] 0.747 0.779 0.759 N4-fields [9] 0.753 0.769 0.784

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.1123

22

MCG [2] 0.747 0.779 0.759 N4-fields [9] 0.753 0.769 0.784 DeepEdge 0.753 0.772 0.807 Table 1: Edge detection results on the BSDS500 benchmark. Our DeepEdge method achieves state-of-the-art contour detections results according to both F-score and AP metrics. Scale ODS OIS AP 64 0.71 0.73 0.76 128 0.72 0.74 0.78 196 0.71 0.73 0.76 Full Image 0.67 0.69 0.57 64, 128 0.72 0.75 0.78 64, 128,196 0.72 0.75 0.78 64,128,196,Full Image 0.75 0.77 0.81 Table 2: Results illustrating the effect of using a multi-scale architecture. Considering multiple scales for contour detection yields significantly higher accuracy relative to a single scale approach. average precisions. Thus, these results suggest that a multiscale approach is highly advantageous in comparison to a single scale setting. 4.3. Advantages of Higher-Level Features In this section, we examine the validity of our earlier

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.1123

23

single scale setting. 4.3. Advantages of Higher-Level Features In this section, we examine the validity of our earlier claim that higher-level object-features enhance contour detection accuracy. In Table 3, we present individual contour detection results using features from the different convolutional layers of KNet . Note that the 4thconvolutional layer produces the most effective features when considering one layer at a time. From our earlier discussion we know that the4thconvolutional layer encodes higher-level object information related to shape and specific object parts. This indicates that object specific cues are particularly beneficial for contour detection accuracy. We also observe that by incorporating features from all Input Image Canny Edges Raw DeepEdges Thresholded DeepEdges Ground Truth Edges Figure 6: Qualitative results produced by our method. Notice how our method learns to distinguish between strong and weak contours. For instance, in the last row of predictions, contours corresponding to zebra stripes are assigned much lower probabilities than contours that correspond to the actual object boundaries separating the zebras from the background. Conv. Layers ODS OIS AP 1st0.66 0.68 0.69 2nd0.71 0.74 0.76

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.1123

24

Conv. Layers ODS OIS AP 1st0.66 0.68 0.69 2nd0.71 0.74 0.76 3rd0.74 0.75 0.79 4th0.74 0.76 0.79 5th0.73 0.74 0.77 All 0.75 0.77 0.81 Table 3: This table shows the advantage of using higherlevel features from the KNet convolutional layers. Individually, the 4thconvolutional layer produces the best contour prediction results, which implies that higher-level object information is indeed beneficial for contour detection. Combining the features from all convolutional layers leads to state-of-the-art results. the convolutional layers, our method achieves state-of-theart contour detection results. This suggests that the features computed by different layers are complementary and that considering information from the entire hierarchy is advantageous.4.4. Pooling Scheme When presenting the architecture of our model, we discussed three different types of pooling: max,average , and center pooling. These three techniques were used to pool the values from the sub-volumes extracted around the center point in each convolutional filter as illustrated in Fig. 5. We now show how each type of pooling affects contour

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.1123

25

We now show how each type of pooling affects contour detection results. Table 4 illustrates that, individually, center pooling yields the best contour detection results. This is expected because the candidate point for which we are trying to predict a contour probability is located at the center of the input patch. However, we note that combining all three types of pooling, achieves better contour detection results than any single pooling technique alone. 4.5. Bifurcation and Training Objective Next, we want to show that that the two independentlytrained classification and regression branches in the bifurcated sub-network provide complementary information that yields improved contour detection accuracy. In Table 5, we present contour detection results achieved by using predictions from the individual branches of the bifurcated subnetwork. Pooling Type ODS OIS AP Average 0.73 0.75 0.78 Max 0.69 0.72 0.73 Center 0.74 0.76 0.8 Avg+Max+Cen 0.75 0.77 0.81 Table 4: Effect of different pooling schemes on contour detection results. Center pooling produces better results than max or average pooling. Combining all three types of pooling further improves the results. Branch ODS OIS AP

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.1123

26

max or average pooling. Combining all three types of pooling further improves the results. Branch ODS OIS AP Classification 0.75 0.76 0.78 Regression 0.74 0.76 0.80 Classification+Regression 0.75 0.77 0.81 Table 5: Contour detection accuracy of the two branches in our bifurcated sub-network. The classification branch yields solid F-score results whereas the regression branch achieves high average precision. Averaging the outputs from these two branches further improve the results. From these results, we observe that using predictions just from the classification branch produces high F-score whereas using predictions only from the regression branch yields high average precision. Combining the predictions from both branches improves the results according to both metrics thus, supporting our claim that separately optimizing edge classification and regression objectives is beneficial to contour detection. 4.6. Qualitative Results Finally, we present qualitative results produced by our method. In Figure 6 we show for each input image example, the set of candidate points produced by the Canny edge detector, the un-thresholded predictions of our method, the

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.1123

27

method. In Figure 6 we show for each input image example, the set of candidate points produced by the Canny edge detector, the un-thresholded predictions of our method, the thresholded predictions, and the ground truth contour map computed as an average of the multiple manual annotations. To generate the thresholded predictions, we use a probability threshold of 0:5. Note that our method successfully distinguishes between strong and weak contours. Specifically, observe that in the last row of Figure 6, our method assigns lower probability to contours corresponding to zebra stripes compared to the contours of the actual object boundary separating the zebras from the background. Thus, in the thresholded version of the prediction, the weak contours inside the zebra bodies are removed and we obtain contour predictions that look very similar to the ground truth. Due to locality of our method, it may be beneficial to apply spectral methods [21, 30] or conditional randomfields [27] on top of our method to further improve its performance. 4.7. Computational Cost In its current form, our method requires about 60KKNet evaluations ( 15Kper scale) to extract the features. Based on the runtimes reported in [14], if executed on a GPU our

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.1123

28

evaluations ( 15Kper scale) to extract the features. Based on the runtimes reported in [14], if executed on a GPU our method would take about 5 minutes and could be made faster using the approach described in [12]. An alternative way to dramatically reduce the runtime of DeepEdge is to interpolate the entries of the feature maps produced by applying the KNet to the full image rather than to individual patches. Such an approach would reduce the number of CNN evaluations needed from 60Kto4(one for each scale), which would allow our method to run in real time even on CPUs. We note that interpolation of features in deep layers has been used successfully in several recent vision papers [29, 11, 19]. Thus, we believe that such an approach could yield nearly equivalent contour detection accuracy, up to a small possible degradation caused by interpolation. Since in this work we were primarily interested in studying the effective advantage enabled by object-level features in contour detection, we have not invested any effort in optimizing the implementation of our method. This will be one of our immediate goals in the future. 5. Conclusions In this work, we presented a multi-scale bifurcated deep network for top-down contour detection. In the past, contour detection has been approached as a bottom-up task

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.1123

29

network for top-down contour detection. In the past, contour detection has been approached as a bottom-up task where low-level features are engineered first, then contour detection is performed, and finally contours may be used as cues for object detection. However, due to a close relationship between object and contour detection tasks, we proposed to invert this pipeline and perform contour detection in a top-down fashion. We demonstrated how to use higher-level object cues to predict contours and showed that considering higher-level object-features leads to a substantial gain in contour detection accuracy. Additionally, we demonstrated that our multi-scale architecture is beneficial to contour prediction as well. By considering multiple scales, our method incorporates local and global information around the candidate contour points, which leads to significantly better contour detection results. Furthermore, we showed that independent optimization of contour classification and regression objectives improves contour prediction accuracy as well. As our experiments indicate, DeepEdge achieves higher average precision results compared to any prior or concurrent work. In conclusion, our results suggest that pure CNN systems can be applied successfully to contour detection and possibly to many other low-level vision tasks. 6. Acknowledgements

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.1123

30

In conclusion, our results suggest that pure CNN systems can be applied successfully to contour detection and possibly to many other low-level vision tasks. 6. Acknowledgements We thank Piotr Teterwak, Du Tran and Mohammad Haris Baig for helpful discussions. This research was funded in part by NSF award CNS-1205521. References [1] P. Arbelaez, M. Maire, C. Fowlkes, and J. Malik. Contour detection and hierarchical image segmentation. IEEE Trans. Pattern Anal. Mach. Intell. , 33(5):898–916, May 2011. 1, 2, 6 [2] P. Arbelaez, J. Pont-Tuset, J. Barron, F. Marqu ´es, and J. Malik. Multiscale combinatorial grouping. In Computer Vision and Pattern Recognition (CVPR) , 2014. 2, 6 [3] E. Borenstein. Combining top-down and bottom-up segmentation. In In Proceedings IEEE workshop on Perceptual Organization in Computer Vision, CVPR , page 46, 2004. 1 [4] C. J. C. Burges. A tutorial on support vector machines for pattern recognition. Data Mining and Knowledge Discovery , 2:121–167, 1998. 2

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.1123

31

[4] C. J. C. Burges. A tutorial on support vector machines for pattern recognition. Data Mining and Knowledge Discovery , 2:121–167, 1998. 2 [5] J. Canny. A computational approach to edge detection. IEEE Trans. Pattern Anal. Mach. Intell. , 8(6):679–698, June 1986. 3 [6] D. Comaniciu, P. Meer, and S. Member. Mean shift: A robust approach toward feature space analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence , 24:603–619, 2002. 6 [7] P. Doll ´ar and C. L. Zitnick. Fast edge detection using structured forests. PAMI , 2015. 2, 6 [8] P. F. Felzenszwalb and D. P. Huttenlocher. Efficient graph-based image segmentation. Int. J. Comput. Vision , 59(2):167–181, Sept. 2004. 6 [9] Y . Ganin and V . S. Lempitsky. N4-fields: Neural network nearest neighbor fields for image transforms. ACCV , 2014. 2, 6

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.1123

32

nearest neighbor fields for image transforms. ACCV , 2014. 2, 6 [10] R. Girshick, J. Donahue, T. Darrell, and J. Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2014. 3 [11] B. Hariharan, P. A. Arbel ´aez, R. B. Girshick, and J. Malik. Hypercolumns for object segmentation and fine-grained localization. CoRR , abs/1411.5752, 2014. 8 [12] F. N. Iandola, M. W. Moskewicz, S. Karayev, R. B. Girshick, T. Darrell, and K. Keutzer. Densenet: Implementing efficient convnet descriptor pyramids. CoRR , abs/1404.1869, 2014. 8 [13] P. Isola, D. Zoran, D. Krishnan, and E. H. Adelson. Crisp boundary detection using pointwise mutual information. In ECCV , 2014. 2, 6

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.1123

33

boundary detection using pointwise mutual information. In ECCV , 2014. 2, 6 [14] Y . Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S. Guadarrama, and T. Darrell. Caffe: Convolutional architecture for fast feature embedding. arXiv preprint arXiv:1408.5093 , 2014. 5, 8 [15] S. Karayev, A. Hertzmann, H. Winnemoeller, A. Agarwala, and T. Darrell. Recognizing image style. CoRR , abs/1311.3715, 2013. 3[16] J. J. Kivinen, C. K. Williams, N. Heess, and D. Technologies. Visual boundary prediction: A deep neural prediction network and quality dissection. AISTATS , 1(2):9, 2014. 2, 6 [17] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks. In F. Pereira, C. Burges, L. Bottou, and K. Weinberger, editors, Advances in Neural Information Processing Systems 25 ,

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.1123

34

F. Pereira, C. Burges, L. Bottou, and K. Weinberger, editors, Advances in Neural Information Processing Systems 25 , pages 1097–1105. Curran Associates, Inc., 2012. 2, 3 [18] J. Lim, C. L. Zitnick, and P. Doll ´ar. Sketch tokens: A learned mid-level representation for contour and object detection. In CVPR , 2013. 2, 6 [19] J. Long, E. Shelhamer, and T. Darrell. Fully convolutional networks for semantic segmentation. CoRR , abs/1411.4038, 2014. 8 [20] M. Maire, S. X. Yu, and P. Perona. Reconstructive sparse code transfer for contour detection and semantic labeling. In Asian Conference on Computer Vision (ACCV) , 2014. 2, 6 [21] J. Malik, S. Belongie, T. Leung, and J. Shi. Contour and texture analysis for image segmentation. Int. J. Comput. Vision , 43(1):7–27, June 2001. 8 [22] T. Malisiewicz, A. Gupta, and A. A. Efros. Ensemble of exemplar-svms for object detection and beyond. In ICCV ,

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.1123

35

exemplar-svms for object detection and beyond. In ICCV , 2011. 5 [23] D. Martin, C. Fowlkes, D. Tal, and J. Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proc. 8th Int’l Conf. Computer Vision , volume 2, pages 416–423, July 2001. 6 [24] A. Opelt and A. Zisserman. A boundary-fragment-model for object detection. In In ECCV , pages 575–588, 2006. 1 [25] P. H. O. Pinheiro and R. Collobert. Recurrent convolutional neural networks for scene parsing. CoRR , abs/1306.2795, 2013. 2 [26] X. Ren and L. Bo. Discriminatively Trained Sparse Code Gradients for Contour Detection. In Advances in Neural Information Processing Systems , December 2012. 2, 6 [27] X. Ren, C. C. Fowlkes, and J. Malik. Scale-invariant contour completion using condition random fields. Technical report. 8

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.1123

36

completion using condition random fields. Technical report. 8 [28] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, A. C. Berg, and L. Fei-Fei. ImageNet Large Scale Visual Recognition Challenge, 2014. 3 [29] P. Sermanet, D. Eigen, X. Zhang, M. Mathieu, R. Fergus, and Y . LeCun. Overfeat: Integrated recognition, localization and detection using convolutional networks. CoRR , abs/1312.6229, 2013. 8 [30] J. Shi and J. Malik. Normalized cuts and image segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence , 22:888–905, 1997. 2, 6, 8 [31] J. Shotton. Contour-based learning for object detection. In In Proc. ICCV , pages 503–510, 2005. 1 [32] Y . Taigman, M. Yang, M. Ranzato, and L. Wolf. Deepface:

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.1123

37

[32] Y . Taigman, M. Yang, M. Ranzato, and L. Wolf. Deepface: Closing the gap to human-level performance in face verification. In Conference on Computer Vision and Pattern Recognition (CVPR) , 2014. 2 [33] A. Toshev and C. Szegedy. Deeppose: Human pose estimation via deep neural networks. CoRR , abs/1312.4659, 2013. 2 [34] N. Zhang, M. Paluri, M. Ranzato, T. Darrell, and L. Bourdev. Panda: Pose aligned networks for deep attribute modeling. CoRR , abs/1311.5591, 2013. 3

1412.1123

DeepEdge: A Multi-Scale Bifurcated Deep Network for Top-Down Contour Detection

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection. We achieve this goal by means of a multi-scale deep network that consists of five convolutional layers and a bifurcated fully-connected sub-network. The section from the input layer to the fifth convolutional layer is fixed and directly lifted from a pre-trained network optimized over a large-scale object classification task. This section of the network is applied to four different scales of the image input. These four parallel and identical streams are then attached to a bifurcated sub-network consisting of two independently-trained branches. One branch learns to predict the contour likelihood (with a classification objective) whereas the other branch is trained to learn the fraction of human labelers agreeing about the contour presence at a given point (with a regression criterion). We show that without any feature engineering our multi-scale deep learning approach achieves state-of-the-art results in contour detection.

http://arxiv.org/pdf/1412.1123

Accepted to CVPR 2015

cs.CV

20141202

20150423

1412.7062

0

Published as a conference paper at ICLR 2015 SEMANTIC IMAGE SEGMENTATION WITH DEEP CONVOLUTIONAL NETS AND FULLY CONNECTED CRF S Liang-Chieh Chen Univ. of California, Los Angeles lcchen@cs.ucla.edu George Papandreou Google Inc. gpapan@google.com Iasonas Kokkinos CentraleSup ´elec and INRIA iasonas.kokkinos@ecp.fr Kevin Murphy Google Inc. kpmurphy@google.com Alan L. Yuille Univ. of California, Los Angeles yuille@stat.ucla.edu ABSTRACT Deep Convolutional Neural Networks (DCNNs) have recently shown state of the art performance in high level vision tasks, such as image classification and object detection. This work brings together methods from DCNNs and probabilistic graphical models for addressing the task of pixel-level classification (also called ”semantic image segmentation”). We show that responses at the final layer of DCNNs are not sufficiently localized for accurate object segmentation. This is

1412.7062

Semantic Image Segmentation with Deep Convolutional Nets and Fully Connected CRFs

Deep Convolutional Neural Networks (DCNNs) have recently shown state of the art performance in high level vision tasks, such as image classification and object detection. This work brings together methods from DCNNs and probabilistic graphical models for addressing the task of pixel-level classification (also called "semantic image segmentation"). We show that responses at the final layer of DCNNs are not sufficiently localized for accurate object segmentation. This is due to the very invariance properties that make DCNNs good for high level tasks. We overcome this poor localization property of deep networks by combining the responses at the final DCNN layer with a fully connected Conditional Random Field (CRF). Qualitatively, our "DeepLab" system is able to localize segment boundaries at a level of accuracy which is beyond previous methods. Quantitatively, our method sets the new state-of-art at the PASCAL VOC-2012 semantic image segmentation task, reaching 71.6% IOU accuracy in the test set. We show how these results can be obtained efficiently: Careful network re-purposing and a novel application of the 'hole' algorithm from the wavelet community allow dense computation of neural net responses at 8 frames per second on a modern GPU.

http://arxiv.org/pdf/1412.7062

14 pages. Updated related work

cs.CV

20141222

20160607

1412.7062

1

DCNNs are not sufficiently localized for accurate object segmentation. This is due to the very invariance properties that make DCNNs good for high level tasks. We overcome this poor localization property of deep networks by combining the responses at the final DCNN layer with a fully connected Conditional Random Field (CRF). Qualitatively, our “DeepLab” system is able to localize segment boundaries at a level of accuracy which is beyond previous methods. Quantitatively, our method sets the new state-of-art at the PASCAL VOC-2012 semantic image segmentation task, reaching 71.6% IOU accuracy in the test set. We show how these results can be obtained efficiently: Careful network re-purposing and a novel application of the ’hole’ algorithm from the wavelet community allow dense computation of neural net responses at 8 frames per second on a modern GPU. 1 I NTRODUCTION Deep Convolutional Neural Networks (DCNNs) had been the method of choice for document recognition since LeCun et al. (1998), but have only recently become the mainstream of high-level vision

1412.7062

Semantic Image Segmentation with Deep Convolutional Nets and Fully Connected CRFs

Deep Convolutional Neural Networks (DCNNs) have recently shown state of the art performance in high level vision tasks, such as image classification and object detection. This work brings together methods from DCNNs and probabilistic graphical models for addressing the task of pixel-level classification (also called "semantic image segmentation"). We show that responses at the final layer of DCNNs are not sufficiently localized for accurate object segmentation. This is due to the very invariance properties that make DCNNs good for high level tasks. We overcome this poor localization property of deep networks by combining the responses at the final DCNN layer with a fully connected Conditional Random Field (CRF). Qualitatively, our "DeepLab" system is able to localize segment boundaries at a level of accuracy which is beyond previous methods. Quantitatively, our method sets the new state-of-art at the PASCAL VOC-2012 semantic image segmentation task, reaching 71.6% IOU accuracy in the test set. We show how these results can be obtained efficiently: Careful network re-purposing and a novel application of the 'hole' algorithm from the wavelet community allow dense computation of neural net responses at 8 frames per second on a modern GPU.

http://arxiv.org/pdf/1412.7062

14 pages. Updated related work

cs.CV

20141222

20160607

1412.7062

2

research. Over the past two years DCNNs have pushed the performance of computer vision systems to soaring heights on a broad array of high-level problems, including image classification (Krizhevsky et al., 2013; Sermanet et al., 2013; Simonyan & Zisserman, 2014; Szegedy et al., 2014; Work initiated when G.P. was with the Toyota Technological Institute at Chicago. The first two authors contributed equally to this work. 1arXiv:1412.7062v4 [cs.CV] 7 Jun 2016 Published as a conference paper at ICLR 2015 Papandreou et al., 2014), object detection (Girshick et al., 2014), fine-grained categorization (Zhang et al., 2014), among others. A common theme in these works is that DCNNs trained in an end-to-end manner deliver strikingly better results than systems relying on carefully engineered representations, such as SIFT or HOG features. This success can be partially attributed to the built-in invariance of DCNNs to local image transformations, which underpins their ability to learn hierarchical abstractions of data (Zeiler & Fergus, 2014). While this invariance is clearly desirable for high-level vision

1412.7062

Semantic Image Segmentation with Deep Convolutional Nets and Fully Connected CRFs

Deep Convolutional Neural Networks (DCNNs) have recently shown state of the art performance in high level vision tasks, such as image classification and object detection. This work brings together methods from DCNNs and probabilistic graphical models for addressing the task of pixel-level classification (also called "semantic image segmentation"). We show that responses at the final layer of DCNNs are not sufficiently localized for accurate object segmentation. This is due to the very invariance properties that make DCNNs good for high level tasks. We overcome this poor localization property of deep networks by combining the responses at the final DCNN layer with a fully connected Conditional Random Field (CRF). Qualitatively, our "DeepLab" system is able to localize segment boundaries at a level of accuracy which is beyond previous methods. Quantitatively, our method sets the new state-of-art at the PASCAL VOC-2012 semantic image segmentation task, reaching 71.6% IOU accuracy in the test set. We show how these results can be obtained efficiently: Careful network re-purposing and a novel application of the 'hole' algorithm from the wavelet community allow dense computation of neural net responses at 8 frames per second on a modern GPU.

http://arxiv.org/pdf/1412.7062

14 pages. Updated related work

cs.CV

20141222

20160607

1412.7062

3

tasks, it can hamper low-level tasks, such as pose estimation (Chen & Yuille, 2014; Tompson et al., 2014) and semantic segmentation - where we want precise localization, rather than abstraction of spatial details. There are two technical hurdles in the application of DCNNs to image labeling tasks: signal downsampling, and spatial ‘insensitivity’ (invariance). The first problem relates to the reduction of signal resolution incurred by the repeated combination of max-pooling and downsampling (‘striding’) performed at every layer of standard DCNNs (Krizhevsky et al., 2013; Simonyan & Zisserman, 2014; Szegedy et al., 2014). Instead, as in Papandreou et al. (2014), we employ the ‘atrous’ (with holes) algorithm originally developed for efficiently computing the undecimated discrete wavelet transform (Mallat, 1999). This allows efficient dense computation of DCNN responses in a scheme substantially simpler than earlier solutions to this problem (Giusti et al., 2013; Sermanet et al., 2013).

1412.7062

Semantic Image Segmentation with Deep Convolutional Nets and Fully Connected CRFs

Deep Convolutional Neural Networks (DCNNs) have recently shown state of the art performance in high level vision tasks, such as image classification and object detection. This work brings together methods from DCNNs and probabilistic graphical models for addressing the task of pixel-level classification (also called "semantic image segmentation"). We show that responses at the final layer of DCNNs are not sufficiently localized for accurate object segmentation. This is due to the very invariance properties that make DCNNs good for high level tasks. We overcome this poor localization property of deep networks by combining the responses at the final DCNN layer with a fully connected Conditional Random Field (CRF). Qualitatively, our "DeepLab" system is able to localize segment boundaries at a level of accuracy which is beyond previous methods. Quantitatively, our method sets the new state-of-art at the PASCAL VOC-2012 semantic image segmentation task, reaching 71.6% IOU accuracy in the test set. We show how these results can be obtained efficiently: Careful network re-purposing and a novel application of the 'hole' algorithm from the wavelet community allow dense computation of neural net responses at 8 frames per second on a modern GPU.

http://arxiv.org/pdf/1412.7062

14 pages. Updated related work

cs.CV

20141222

20160607

1412.7062

4

The second problem relates to the fact that obtaining object-centric decisions from a classifier requires invariance to spatial transformations, inherently limiting the spatial accuracy of the DCNN model. We boost our model’s ability to capture fine details by employing a fully-connected Conditional Random Field (CRF). Conditional Random Fields have been broadly used in semantic segmentation to combine class scores computed by multi-way classifiers with the low-level information captured by the local interactions of pixels and edges (Rother et al., 2004; Shotton et al., 2009) or superpixels (Lucchi et al., 2011). Even though works of increased sophistication have been proposed to model the hierarchical dependency (He et al., 2004; Ladicky et al., 2009; Lempitsky et al., 2011) and/or high-order dependencies of segments (Delong et al., 2012; Gonfaus et al., 2010; Kohli et al., 2009; Chen et al., 2013; Wang et al., 2015), we use the fully connected pairwise CRF proposed by Kr¨ahenb ¨uhl & Koltun (2011) for its efficient computation, and ability to capture fine edge details

1412.7062

Semantic Image Segmentation with Deep Convolutional Nets and Fully Connected CRFs

Deep Convolutional Neural Networks (DCNNs) have recently shown state of the art performance in high level vision tasks, such as image classification and object detection. This work brings together methods from DCNNs and probabilistic graphical models for addressing the task of pixel-level classification (also called "semantic image segmentation"). We show that responses at the final layer of DCNNs are not sufficiently localized for accurate object segmentation. This is due to the very invariance properties that make DCNNs good for high level tasks. We overcome this poor localization property of deep networks by combining the responses at the final DCNN layer with a fully connected Conditional Random Field (CRF). Qualitatively, our "DeepLab" system is able to localize segment boundaries at a level of accuracy which is beyond previous methods. Quantitatively, our method sets the new state-of-art at the PASCAL VOC-2012 semantic image segmentation task, reaching 71.6% IOU accuracy in the test set. We show how these results can be obtained efficiently: Careful network re-purposing and a novel application of the 'hole' algorithm from the wavelet community allow dense computation of neural net responses at 8 frames per second on a modern GPU.

http://arxiv.org/pdf/1412.7062

14 pages. Updated related work

cs.CV

20141222

20160607