Datasets:

jamescalam

/

llama-2-arxiv-papers-chunked

like 17

1412.7062

5

Kr¨ahenb ¨uhl & Koltun (2011) for its efficient computation, and ability to capture fine edge details while also catering for long range dependencies. That model was shown in Kr ¨ahenb ¨uhl & Koltun (2011) to largely improve the performance of a boosting-based pixel-level classifier, and in our work we demonstrate that it leads to state-of-the-art results when coupled with a DCNN-based pixel-level classifier. The three main advantages of our “DeepLab” system are (i) speed: by virtue of the ‘atrous’ algorithm, our dense DCNN operates at 8 fps, while Mean Field Inference for the fully-connected CRF requires 0.5 second, (ii) accuracy: we obtain state-of-the-art results on the PASCAL semantic segmentation challenge, outperforming the second-best approach of Mostajabi et al. (2014) by a margin of 7.2 %and (iii) simplicity: our system is composed of a cascade of two fairly well-established modules, DCNNs and CRFs. 2 R ELATED WORK

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

6

of 7.2 %and (iii) simplicity: our system is composed of a cascade of two fairly well-established modules, DCNNs and CRFs. 2 R ELATED WORK Our system works directly on the pixel representation, similarly to Long et al. (2014). This is in contrast to the two-stage approaches that are now most common in semantic segmentation with DCNNs: such techniques typically use a cascade of bottom-up image segmentation and DCNN-based region classification, which makes the system commit to potential errors of the front-end segmentation system. For instance, the bounding box proposals and masked regions delivered by (Arbel ´aez et al., 2014; Uijlings et al., 2013) are used in Girshick et al. (2014) and (Hariharan et al., 2014b) as inputs to a DCNN to introduce shape information into the classification process. Similarly, the authors of Mostajabi et al. (2014) rely on a superpixel representation. A celebrated non-DCNN precursor to these works is the second order pooling method of (Carreira et al., 2012) which also assigns labels to the regions proposals delivered by (Carreira & Sminchisescu, 2012). Understanding the perils

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

7

to the regions proposals delivered by (Carreira & Sminchisescu, 2012). Understanding the perils of committing to a single segmentation, the authors of Cogswell et al. (2014) build on (Yadollahpour et al., 2013) to explore a diverse set of CRF-based segmentation proposals, computed also by (Carreira & Sminchisescu, 2012). These segmentation proposals are then re-ranked according to a DCNN trained in particular for this reranking task. Even though this approach explicitly tries to handle the temperamental nature of a front-end segmentation algorithm, there is still no explicit ex2 Published as a conference paper at ICLR 2015 ploitation of the DCNN scores in the CRF-based segmentation algorithm: the DCNN is only applied post-hoc, while it would make sense to directly try to use its results during segmentation. Moving towards works that lie closer to our approach, several other researchers have considered the use of convolutionally computed DCNN features for dense image labeling. Among the first have been Farabet et al. (2013) who apply DCNNs at multiple image resolutions and then employ a segmentation tree to smooth the prediction results; more recently, Hariharan et al. (2014a) propose to

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

8

segmentation tree to smooth the prediction results; more recently, Hariharan et al. (2014a) propose to concatenate the computed inter-mediate feature maps within the DCNNs for pixel classification, and Dai et al. (2014) propose to pool the inter-mediate feature maps by region proposals. Even though these works still employ segmentation algorithms that are decoupled from the DCNN classifier’s results, we believe it is advantageous that segmentation is only used at a later stage, avoiding the commitment to premature decisions. More recently, the segmentation-free techniques of (Long et al., 2014; Eigen & Fergus, 2014) directly apply DCNNs to the whole image in a sliding window fashion, replacing the last fully connected layers of a DCNN by convolutional layers. In order to deal with the spatial localization issues outlined in the beginning of the introduction, Long et al. (2014) upsample and concatenate the scores from inter-mediate feature maps, while Eigen & Fergus (2014) refine the prediction result from coarse to fine by propagating the coarse results to another DCNN. The main difference between our model and other state-of-the-art models is the combination of

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

9

from coarse to fine by propagating the coarse results to another DCNN. The main difference between our model and other state-of-the-art models is the combination of pixel-level CRFs and DCNN-based ‘unary terms’. Focusing on the closest works in this direction, Cogswell et al. (2014) use CRFs as a proposal mechanism for a DCNN-based reranking system, while Farabet et al. (2013) treat superpixels as nodes for a local pairwise CRF and use graph-cuts for discrete inference; as such their results can be limited by errors in superpixel computations, while ignoring long-range superpixel dependencies. Our approach instead treats every pixel as a CRF node, exploits long-range dependencies, and uses CRF inference to directly optimize a DCNN-driven cost function. We note that mean field had been extensively studied for traditional image segmentation/edge detection tasks, e.g., (Geiger & Girosi, 1991; Geiger & Yuille, 1991; Kokkinos et al., 2008), but recently Kr ¨ahenb ¨uhl & Koltun (2011) showed that the inference can be very efficient for fully connected CRF and particularly effective in the context of semantic segmentation.

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

10

fully connected CRF and particularly effective in the context of semantic segmentation. After the first version of our manuscript was made publicly available, it came to our attention that two other groups have independently and concurrently pursued a very similar direction, combining DCNNs and densely connected CRFs (Bell et al., 2014; Zheng et al., 2015). There are several differences in technical aspects of the respective models. Bell et al. (2014) focus on the problem of material classification, while Zheng et al. (2015) unroll the CRF mean-field inference steps to convert the whole system into an end-to-end trainable feed-forward network. We have updated our proposed “DeepLab” system with much improved methods and results in our latest work (Chen et al., 2016). We refer the interested reader to the paper for details. 3 C ONVOLUTIONAL NEURAL NETWORKS FOR DENSE IMAGE LABELING Herein we describe how we have re-purposed and finetuned the publicly available Imagenetpretrained state-of-art 16-layer classification network of (Simonyan & Zisserman, 2014) (VGG-16) into an efficient and effective dense feature extractor for our dense semantic image segmentation

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

11

into an efficient and effective dense feature extractor for our dense semantic image segmentation system. 3.1 E FFICIENT DENSE SLIDING WINDOW FEATURE EXTRACTION WITH THE HOLE ALGORITHM Dense spatial score evaluation is instrumental in the success of our dense CNN feature extractor. As a first step to implement this, we convert the fully-connected layers of VGG-16 into convolutional ones and run the network in a convolutional fashion on the image at its original resolution. However this is not enough as it yields very sparsely computed detection scores (with a stride of 32 pixels). To compute scores more densely at our target stride of 8 pixels, we develop a variation of the method previously employed by Giusti et al. (2013); Sermanet et al. (2013). We skip subsampling after the last two max-pooling layers in the network of Simonyan & Zisserman (2014) and modify the convolutional filters in the layers that follow them by introducing zeros to increase their length ( 2in 3 Published as a conference paper at ICLR 2015 Input stride Output stride Figure 1: Illustration of the hole algorithm in 1-D, when kernel size = 3 ,input stride = 2 , and

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

12

Input stride Output stride Figure 1: Illustration of the hole algorithm in 1-D, when kernel size = 3 ,input stride = 2 , and output stride = 1 . the last three convolutional layers and 4in the first fully connected layer). We can implement this more efficiently by keeping the filters intact and instead sparsely sample the feature maps on which they are applied on using an input stride of 2 or 4 pixels, respectively. This approach, illustrated in Fig. 1 is known as the ‘hole algorithm’ (‘atrous algorithm’) and has been developed before for efficient computation of the undecimated wavelet transform (Mallat, 1999). We have implemented this within the Caffe framework (Jia et al., 2014) by adding to the im2col function (it converts multichannel feature maps to vectorized patches) the option to sparsely sample the underlying feature map. This approach is generally applicable and allows us to efficiently compute dense CNN feature maps at any target subsampling rate without introducing any approximations. We finetune the model weights of the Imagenet-pretrained VGG-16 network to adapt it to the image

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

13

We finetune the model weights of the Imagenet-pretrained VGG-16 network to adapt it to the image classification task in a straightforward fashion, following the procedure of Long et al. (2014). We replace the 1000-way Imagenet classifier in the last layer of VGG-16 with a 21-way one. Our loss function is the sum of cross-entropy terms for each spatial position in the CNN output map (subsampled by 8 compared to the original image). All positions and labels are equally weighted in the overall loss function. Our targets are the ground truth labels (subsampled by 8). We optimize the objective function with respect to the weights at all network layers by the standard SGD procedure of Krizhevsky et al. (2013). During testing, we need class score maps at the original image resolution. As illustrated in Figure 2 and further elaborated in Section 4.1, the class score maps (corresponding to log-probabilities) are quite smooth, which allows us to use simple bilinear interpolation to increase their resolution by a factor of 8 at a negligible computational cost. Note that the method of Long et al. (2014) does not use the hole algorithm and produces very coarse scores (subsampled by a factor of 32) at the CNN

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

14

use the hole algorithm and produces very coarse scores (subsampled by a factor of 32) at the CNN output. This forced them to use learned upsampling layers, significantly increasing the complexity and training time of their system: Fine-tuning our network on PASCAL VOC 2012 takes about 10 hours, while they report a training time of several days (both timings on a modern GPU). 3.2 C ONTROLLING THE RECEPTIVE FIELD SIZE AND ACCELERATING DENSE COMPUTATION WITH CONVOLUTIONAL NETS Another key ingredient in re-purposing our network for dense score computation is explicitly controlling the network’s receptive field size. Most recent DCNN-based image recognition methods rely on networks pre-trained on the Imagenet large-scale classification task. These networks typically have large receptive field size: in the case of the VGG-16 net we consider, its receptive field is224224(with zero-padding) and 404404pixels if the net is applied convolutionally. After

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

15

is224224(with zero-padding) and 404404pixels if the net is applied convolutionally. After converting the network to a fully convolutional one, the first fully connected layer has 4,096 filters of large 77spatial size and becomes the computational bottleneck in our dense score map computation. We have addressed this practical problem by spatially subsampling (by simple decimation) the first FC layer to 44(or33) spatial size. This has reduced the receptive field of the network down to 128128(with zero-padding) or 308308(in convolutional mode) and has reduced computation time for the first FC layer by 23times. Using our Caffe-based implementation and a Titan GPU, the resulting VGG-derived network is very efficient: Given a 306306input image, it produces 3939 4 Published as a conference paper at ICLR 2015 dense raw feature scores at the top of the network at a rate of about 8 frames/sec during testing. The speed during training is 3 frames/sec. We have also successfully experimented with reducing the number of channels at the fully connected layers from 4,096 down to 1,024, considerably further

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

16

speed during training is 3 frames/sec. We have also successfully experimented with reducing the number of channels at the fully connected layers from 4,096 down to 1,024, considerably further decreasing computation time and memory footprint without sacrificing performance, as detailed in Section 5. Using smaller networks such as Krizhevsky et al. (2013) could allow video-rate test-time dense feature computation even on light-weight GPUs. 4 D ETAILED BOUNDARY RECOVERY : FULLY -CONNECTED CONDITIONAL RANDOM FIELDS AND MULTI -SCALE PREDICTION 4.1 D EEPCONVOLUTIONAL NETWORKS AND THE LOCALIZATION CHALLENGE As illustrated in Figure 2, DCNN score maps can reliably predict the presence and rough position of objects in an image but are less well suited for pin-pointing their exact outline. There is a natural trade-off between classification accuracy and localization accuracy with convolutional networks: Deeper models with multiple max-pooling layers have proven most successful in classification tasks, however their increased invariance and large receptive fields make the problem of inferring position from the scores at their top output levels more challenging.

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

17

however their increased invariance and large receptive fields make the problem of inferring position from the scores at their top output levels more challenging. Recent work has pursued two directions to address this localization challenge. The first approach is to harness information from multiple layers in the convolutional network in order to better estimate the object boundaries (Long et al., 2014; Eigen & Fergus, 2014). The second approach is to employ a super-pixel representation, essentially delegating the localization task to a low-level segmentation method. This route is followed by the very successful recent method of Mostajabi et al. (2014). In Section 4.2, we pursue a novel alternative direction based on coupling the recognition capacity of DCNNs and the fine-grained localization accuracy of fully connected CRFs and show that it is remarkably successful in addressing the localization challenge, producing accurate semantic segmentation results and recovering object boundaries at a level of detail that is well beyond the reach of existing methods. 4.2 F ULLY -CONNECTED CONDITIONAL RANDOM FIELDS FOR ACCURATE LOCALIZATION Image/G.T. DCNN output CRF Iteration 1 CRF Iteration 2 CRF Iteration 10

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

18

Image/G.T. DCNN output CRF Iteration 1 CRF Iteration 2 CRF Iteration 10 Figure 2: Score map (input before softmax function) and belief map (output of softmax function) for Aeroplane. We show the score (1st row) and belief (2nd row) maps after each mean field iteration. The output of last DCNN layer is used as input to the mean field inference. Best viewed in color. Traditionally, conditional random fields (CRFs) have been employed to smooth noisy segmentation maps (Rother et al., 2004; Kohli et al., 2009). Typically these models contain energy terms that couple neighboring nodes, favoring same-label assignments to spatially proximal pixels. Qualitatively, the primary function of these short-range CRFs has been to clean up the spurious predictions of weak classifiers built on top of local hand-engineered features. Compared to these weaker classifiers, modern DCNN architectures such as the one we use in this work produce score maps and semantic label predictions which are qualitatively different. As illustrated in Figure 2, the score maps are typically quite smooth and produce homogeneous classification

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

19

work produce score maps and semantic label predictions which are qualitatively different. As illustrated in Figure 2, the score maps are typically quite smooth and produce homogeneous classification results. In this regime, using short-range CRFs can be detrimental, as our goal should be to recover detailed local structure rather than further smooth it. Using contrast-sensitive potentials (Rother 5 Published as a conference paper at ICLR 2015 Deep Convolutional Neural Network InputAeroplane Coarse Score map Bi-linear Interpolation Fully Connected CRF Final Output Figure 3: Model Illustration. The coarse score map from Deep Convolutional Neural Network (with fully convolutional layers) is upsampled by bi-linear interpolation. A fully connected CRF is applied to refine the segmentation result. Best viewed in color. et al., 2004) in conjunction to local-range CRFs can potentially improve localization but still miss thin-structures and typically requires solving an expensive discrete optimization problem. To overcome these limitations of short-range CRFs, we integrate into our system the fully connected CRF model of Kr ¨ahenb ¨uhl & Koltun (2011). The model employs the energy function E(x) =X ii(xi) +X

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

20

E(x) =X ii(xi) +X ijij(xi;xj) (1) where xis the label assignment for pixels. We use as unary potential i(xi) =logP(xi), where P(xi)is the label assignment probability at pixel ias computed by DCNN. The pairwise potential isij(xi;xj) =(xi;xj)PK m=1wm
km(fi;fj), where(xi;xj) = 1 ifxi6=xj, and zero otherwise ( i.e., Potts Model). There is one pairwise term for each pair of pixels iandjin the image no matter how far from each other they lie, i.e. the model’s factor graph is fully connected. Each km is the Gaussian kernel depends on features (denoted as f) extracted for pixel iandjand is weighted by parameter wm. We adopt bilateral position and color terms, specifically, the kernels are w1exp jjpipjjj2 22jjIiIjjj2 22  +w2exp jjpipjjj2 22  (2)

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

21

22  +w2exp jjpipjjj2 22  (2) where the first kernel depends on both pixel positions (denoted as p) and pixel color intensities (denoted asI), and the second kernel only depends on pixel positions. The hyper parameters , and control the “scale” of the Gaussian kernels. Crucially, this model is amenable to efficient approximate probabilistic inference (Kr ¨ahenb ¨uhl & Koltun, 2011). The message passing updates under a fully decomposable mean field approximationb(x) =Q ibi(xi)can be expressed as convolutions with a Gaussian kernel in feature space. High-dimensional filtering algorithms (Adams et al., 2010) significantly speed-up this computation resulting in an algorithm that is very fast in practice, less that 0.5 sec on average for Pascal VOC images using the publicly available implementation of (Kr ¨ahenb ¨uhl & Koltun, 2011). 4.3 M ULTI -SCALE PREDICTION

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

22

images using the publicly available implementation of (Kr ¨ahenb ¨uhl & Koltun, 2011). 4.3 M ULTI -SCALE PREDICTION Following the promising recent results of (Hariharan et al., 2014a; Long et al., 2014) we have also explored a multi-scale prediction method to increase the boundary localization accuracy. Specifically, we attach to the input image and the output of each of the first four max pooling layers a two-layer MLP (first layer: 128 3x3 convolutional filters, second layer: 128 1x1 convolutional filters) whose feature map is concatenated to the main network’s last layer feature map. The aggregate feature map fed into the softmax layer is thus enhanced by 5 * 128 = 640 channels. We only adjust the newly added weights, keeping the other network parameters to the values learned by the method of Section 3. As discussed in the experimental section, introducing these extra direct connections from fine-resolution layers improves localization performance, yet the effect is not as dramatic as the one obtained with the fully-connected CRF. 6 Published as a conference paper at ICLR 2015 Method mean IOU (%) DeepLab 59.80 DeepLab-CRF 63.74

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

23

6 Published as a conference paper at ICLR 2015 Method mean IOU (%) DeepLab 59.80 DeepLab-CRF 63.74 DeepLab-MSc 61.30 DeepLab-MSc-CRF 65.21 DeepLab-7x7 64.38 DeepLab-CRF-7x7 67.64 DeepLab-LargeFOV 62.25 DeepLab-CRF-LargeFOV 67.64 DeepLab-MSc-LargeFOV 64.21 DeepLab-MSc-CRF-LargeFOV 68.70Method mean IOU (%) MSRA-CFM 61.8 FCN-8s 62.2 TTI-Zoomout-16 64.4 DeepLab-CRF 66.4 DeepLab-MSc-CRF 67.1 DeepLab-CRF-7x7 70.3 DeepLab-CRF-LargeFOV 70.3 DeepLab-MSc-CRF-LargeFOV 71.6 (a) (b) Table 1: (a) Performance of our proposed models on the PASCAL VOC 2012 ‘val’ set (with training in the augmented ‘train’ set). The best performance is achieved by exploiting both multi-scale

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

24

in the augmented ‘train’ set). The best performance is achieved by exploiting both multi-scale features and large field-of-view. (b) Performance of our proposed models (with training in the augmented ‘trainval’ set) compared to other state-of-art methods on the PASCAL VOC 2012 ‘test’ set. 5 E XPERIMENTAL EVALUATION Dataset We test our DeepLab model on the PASCAL VOC 2012 segmentation benchmark (Everingham et al., 2014), consisting of 20 foreground object classes and one background class. The original dataset contains 1;464,1;449, and 1;456images for training, validation, and testing, respectively. The dataset is augmented by the extra annotations provided by Hariharan et al. (2011), resulting in 10;582training images. The performance is measured in terms of pixel intersectionover-union (IOU) averaged across the 21 classes. Training We adopt the simplest form of piecewise training, decoupling the DCNN and CRF training stages, assuming the unary terms provided by the DCNN are fixed during CRF training. For DCNN training we employ the VGG-16 network which has been pre-trained on ImageNet. We

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

25

For DCNN training we employ the VGG-16 network which has been pre-trained on ImageNet. We fine-tuned the VGG-16 network on the VOC 21-way pixel-classification task by stochastic gradient descent on the cross-entropy loss function, as described in Section 3.1. We use a mini-batch of 20 images and initial learning rate of 0:001(0:01for the final classifier layer), multiplying the learning rate by 0.1 at every 2000 iterations. We use momentum of 0:9and a weight decay of 0:0005 . After the DCNN has been fine-tuned, we cross-validate the parameters of the fully connected CRF model in Eq. (2) along the lines of Kr ¨ahenb ¨uhl & Koltun (2011). We use the default values of w2= 3 and = 3 and we search for the best values of w1,, andby cross-validation on a small subset of the validation set (we use 100 images). We employ coarse-to-fine search scheme.

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

26

small subset of the validation set (we use 100 images). We employ coarse-to-fine search scheme. Specifically, the initial search range of the parameters are w12[5;10],2[50 : 10 : 100] and 2[3 : 1 : 10] (MATLAB notation), and then we refine the search step sizes around the first round’s best values. We fix the number of mean field iterations to 10 for all reported experiments. Evaluation on Validation set We conduct the majority of our evaluations on the PASCAL ‘val’ set, training our model on the augmented PASCAL ‘train’ set. As shown in Tab. 1 (a), incorporating the fully connected CRF to our model (denoted by DeepLab-CRF) yields a substantial performance boost, about 4% improvement over DeepLab. We note that the work of Kr ¨ahenb ¨uhl & Koltun (2011) improved the 27:6%result of TextonBoost (Shotton et al., 2009) to 29:1%, which makes the improvement we report here (from 59:8%to63:7%) all the more impressive.

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

27

improvement we report here (from 59:8%to63:7%) all the more impressive. Turning to qualitative results, we provide visual comparisons between DeepLab and DeepLab-CRF in Fig. 7. Employing a fully connected CRF significantly improves the results, allowing the model to accurately capture intricate object boundaries. Multi-Scale features We also exploit the features from the intermediate layers, similar to Hariharan et al. (2014a); Long et al. (2014). As shown in Tab. 1 (a), adding the multi-scale features to our 7 Published as a conference paper at ICLR 2015 Method kernel size input stride receptive field # parameters mean IOU (%) Training speed (img/sec) DeepLab-CRF-7x7 77 4 224 134.3M 67.64 1.44 DeepLab-CRF 44 4 128 65.1M 63.74 2.90 DeepLab-CRF-4x4 44 8 224 65.1M 67.14 2.90 DeepLab-CRF-LargeFOV 33 12 224 20.5M 67.64 4.84 Table 2: Effect of Field-Of-View. We show the performance (after CRF) and training speed on the

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

28

Table 2: Effect of Field-Of-View. We show the performance (after CRF) and training speed on the PASCAL VOC 2012 ‘val’ set as the function of (1) the kernel size of first fully connected layer, (2) the input stride value employed in the atrous algorithm. DeepLab model (denoted as DeepLab-MSc) improves about 1:5%performance, and further incorporating the fully connected CRF (denoted as DeepLab-MSc-CRF) yields about 4% improvement. The qualitative comparisons between DeepLab and DeepLab-MSc are shown in Fig. 4. Leveraging the multi-scale features can slightly refine the object boundaries. Field of View The ‘atrous algorithm’ we employed allows us to arbitrarily control the Field-ofView (FOV) of the models by adjusting the input stride, as illustrated in Fig. 1. In Tab. 2, we experiment with several kernel sizes and input strides at the first fully connected layer. The method, DeepLab-CRF-7x7, is the direct modification from VGG-16 net, where the kernel size = 77and input stride = 4. This model yields performance of 67:64% on the ‘val’ set, but it is relatively slow

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

29

input stride = 4. This model yields performance of 67:64% on the ‘val’ set, but it is relatively slow (1:44images per second during training). We have improved model speed to 2:9images per second by reducing the kernel size to 44. We have experimented with two such network variants with different FOV sizes, DeepLab-CRF and DeepLab-CRF-4x4; the latter has large FOV ( i.e., large input stride) and attains better performance. Finally, we employ kernel size 33and input stride = 12, and further change the filter sizes from 4096 to 1024 for the last two layers. Interestingly, the resulting model, DeepLab-CRF-LargeFOV , matches the performance of the expensive DeepLabCRF-7x7. At the same time, it is 3:36times faster to run and has significantly fewer parameters (20.5M instead of 134.3M). The performance of several model variants is summarized in Tab. 1, showing the benefit of exploiting multi-scale features and large FOV . Figure 4: Incorporating multi-scale features improves the boundary segmentation. We show the

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

30

multi-scale features and large FOV . Figure 4: Incorporating multi-scale features improves the boundary segmentation. We show the results obtained by DeepLab and DeepLab-MSc in the first and second row, respectively. Best viewed in color. Mean Pixel IOU along Object Boundaries To quantify the accuracy of the proposed model near object boundaries, we evaluate the segmentation accuracy with an experiment similar to Kohli et al. (2009); Kr ¨ahenb ¨uhl & Koltun (2011). Specifically, we use the ‘void’ label annotated in val set, which usually occurs around object boundaries. We compute the mean IOU for those pixels that are located within a narrow band (called trimap) of ‘void’ labels. As shown in Fig. 5, exploiting the multi-scale features from the intermediate layers and refining the segmentation results by a fully connected CRF significantly improve the results around object boundaries. Comparison with State-of-art In Fig. 6, we qualitatively compare our proposed model, DeepLabCRF, with two state-of-art models: FCN-8s (Long et al., 2014) and TTI-Zoomout-16 (Mostajabi

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

31

et al., 2014) on the ‘val’ set (the results are extracted from their papers). Our model is able to capture the intricate object boundaries. 8 Published as a conference paper at ICLR 2015 0 510 15 20 25 30 35 405560657075808590Pixelwise Accuracy (%) Trimap Width (pixels) DL−MSc−CRF DeepLab−CRF DeepLab−MSc DeepLab 0 510 15 20 25 30 35 4035404550556065mean IOU (%) Trimap Width (pixels) DL−MSc−CRF DeepLab−CRF DeepLab−MSc DeepLab (a) (b) (c) Figure 5: (a) Some trimap examples (top-left: image. top-right: ground-truth. bottom-left: trimap of 2 pixels. bottom-right: trimap of 10 pixels). Quality of segmentation result within a band around the object boundaries for the proposed methods. (b) Pixelwise accuracy. (c) Pixel mean IOU. (a) FCN-8s vs. DeepLab-CRF (b) TTI-Zoomout-16 vs. DeepLab-CRF

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

32

(a) FCN-8s vs. DeepLab-CRF (b) TTI-Zoomout-16 vs. DeepLab-CRF Figure 6: Comparisons with state-of-the-art models on the val set. First row: images. Second row: ground truths. Third row: other recent models (Left: FCN-8s, Right: TTI-Zoomout-16). Fourth row: our DeepLab-CRF. Best viewed in color. Reproducibility We have implemented the proposed methods by extending the excellent Caffe framework (Jia et al., 2014). We share our source code, configuration files, and trained models that allow reproducing the results in this paper at a companion web site https://bitbucket.org/ deeplab/deeplab-public . Test set results Having set our model choices on the validation set, we evaluate our model variants on the PASCAL VOC 2012 official ‘test’ set. As shown in Tab. 3, our DeepLab-CRF and DeepLabMSc-CRF models achieve performance of 66:4%and67:1%mean IOU1, respectively. Our models

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

33

outperform all the other state-of-the-art models (specifically, TTI-Zoomout-16 (Mostajabi et al., 2014), FCN-8s (Long et al., 2014), and MSRA-CFM (Dai et al., 2014)). When we increase the FOV of the models, DeepLab-CRF-LargeFOV yields performance of 70:3%, the same as DeepLab-CRF7x7, while its training speed is faster. Furthermore, our best model, DeepLab-MSc-CRF-LargeFOV , attains the best performance of 71:6%by employing both multi-scale features and large FOV . 1http://host.robots.ox.ac.uk:8080/leaderboard/displaylb.php? challengeid=11&compid=6 9 Published as a conference paper at ICLR 2015 Figure 7: Visualization results on VOC 2012-val. For each row, we show the input image, the segmentation result delivered by the DCNN (DeepLab), and the refined segmentation result of the Fully Connected CRF (DeepLab-CRF). We show our failure modes in the last three rows. Best viewed in color. 10

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

34

Fully Connected CRF (DeepLab-CRF). We show our failure modes in the last three rows. Best viewed in color. 10 Published as a conference paper at ICLR 2015 Method bkg aero bike bird boat bottle bus car cat chair cow table dog horse mbike person plant sheep sofa train tv mean MSRA-CFM -75.7 26.7 69.5 48.8 65.6 81.0 69.2 73.3 30.0 68.7 51.5 69.1 68.1 71.7 67.5 50.4 66.5 44.4 58.9 53.5 61.8 FCN-8s -76.8 34.2 68.9 49.4 60.3 75.3 74.7 77.6 21.4 62.5 46.8 71.8 63.9 76.5 73.9 45.2 72.4 37.4 70.9 55.1 62.2 TTI-Zoomout-16 89.8 81.9 35.1 78.2 57.4 56.5 80.5 74.0 79.8 22.4 69.6 53.7 74.0 76.0 76.6 68.8 44.3 70.2 40.2 68.9 55.3 64.4

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

35

DeepLab-CRF 92.1 78.4 33.1 78.2 55.6 65.3 81.3 75.5 78.6 25.3 69.2 52.7 75.2 69.0 79.1 77.6 54.7 78.3 45.1 73.3 56.2 66.4 DeepLab-MSc-CRF 92.6 80.4 36.8 77.4 55.2 66.4 81.5 77.5 78.9 27.1 68.2 52.7 74.3 69.6 79.4 79.0 56.9 78.8 45.2 72.7 59.3 67.1 DeepLab-CRF-7x7 92.8 83.9 36.6 77.5 58.4 68.0 84.6 79.7 83.1 29.5 74.6 59.3 78.9 76.0 82.1 80.6 60.3 81.7 49.2 78.0 60.7 70.3 DeepLab-CRF-LargeFOV 92.6 83.5 36.6 82.5 62.3 66.5 85.4 78.5 83.7 30.4 72.9 60.4 78.5 75.5 82.1 79.7 58.2 82.0 48.8 73.7 63.3 70.3

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

36

DeepLab-MSc-CRF-LargeFOV 93.1 84.4 54.5 81.5 63.6 65.9 85.1 79.1 83.4 30.7 74.1 59.8 79.0 76.1 83.2 80.8 59.7 82.2 50.4 73.1 63.7 71.6 Table 3: Labeling IOU (%) on the PASCAL VOC 2012 test set, using the trainval set for training. 6 D ISCUSSION Our work combines ideas from deep convolutional neural networks and fully-connected conditional random fields, yielding a novel method able to produce semantically accurate predictions and detailed segmentation maps, while being computationally efficient. Our experimental results show that the proposed method significantly advances the state-of-art in the challenging PASCAL VOC 2012 semantic image segmentation task. There are multiple aspects in our model that we intend to refine, such as fully integrating its two main components (CNN and CRF) and train the whole system in an end-to-end fashion, similar to Kr¨ahenb ¨uhl & Koltun (2013); Chen et al. (2014); Zheng et al. (2015). We also plan to experiment

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

37

Kr¨ahenb ¨uhl & Koltun (2013); Chen et al. (2014); Zheng et al. (2015). We also plan to experiment with more datasets and apply our method to other sources of data such as depth maps or videos. Recently, we have pursued model training with weakly supervised annotations, in the form of bounding boxes or image-level labels (Papandreou et al., 2015). At a higher level, our work lies in the intersection of convolutional neural networks and probabilistic graphical models. We plan to further investigate the interplay of these two powerful classes of methods and explore their synergistic potential for solving challenging computer vision tasks. ACKNOWLEDGMENTS This work was partly supported by ARO 62250-CS, NIH Grant 5R01EY022247-03, EU Project RECONFIG FP7-ICT-600825 and EU Project MOBOT FP7-ICT-2011-600796. We also gratefully acknowledge the support of NVIDIA Corporation with the donation of GPUs used for this research. We would like to thank the anonymous reviewers for their detailed comments and constructive feedback. PAPER REVISIONS Here we present the list of major paper revisions for the convenience of the readers. v1 Submission to ICLR 2015. Introduces the model DeepLab-CRF, which attains the performance

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

38

v1 Submission to ICLR 2015. Introduces the model DeepLab-CRF, which attains the performance of66:4%on PASCAL VOC 2012 test set. v2 Rebuttal for ICLR 2015. Adds the model DeepLab-MSc-CRF, which incorporates multi-scale features from the intermediate layers. DeepLab-MSc-CRF yields the performance of 67:1%on PASCAL VOC 2012 test set. v3 Camera-ready for ICLR 2015. Experiments with large Field-Of-View. On PASCAL VOC 2012 test set, DeepLab-CRF-LargeFOV achieves the performance of 70:3%. When exploiting both mutliscale features and large FOV , DeepLab-MSc-CRF-LargeFOV attains the performance of 71:6%. v4 Reference to our updated “DeepLab” system (Chen et al., 2016) with much improved results. REFERENCES Adams, A., Baek, J., and Davis, M. A. Fast high-dimensional filtering using the permutohedral lattice. In Computer Graphics Forum , 2010.

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

39

lattice. In Computer Graphics Forum , 2010. Arbel ´aez, P., Pont-Tuset, J., Barron, J. T., Marques, F., and Malik, J. Multiscale combinatorial grouping. In CVPR , 2014. 11 Published as a conference paper at ICLR 2015 Bell, S., Upchurch, P., Snavely, N., and Bala, K. Material recognition in the wild with the materials in context database. arXiv:1412.0623 , 2014. Carreira, J. and Sminchisescu, C. Cpmc: Automatic object segmentation using constrained parametric min-cuts. PAMI , 2012. Carreira, J., Caseiro, R., Batista, J., and Sminchisescu, C. Semantic segmentation with second-order pooling. In ECCV , 2012. Chen, L.-C., Papandreou, G., and Yuille, A. Learning a dictionary of shape epitomes with applications to image labeling. In ICCV , 2013. Chen, L.-C., Schwing, A., Yuille, A., and Urtasun, R. Learning deep structured models. arXiv:1407.2538 , 2014.

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

40

arXiv:1407.2538 , 2014. Chen, L.-C., Papandreou, G., Kokkinos, I., Murphy, K., and Yuille, A. L. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. arXiv:1606.00915 , 2016. Chen, X. and Yuille, A. L. Articulated pose estimation by a graphical model with image dependent pairwise relations. In NIPS , 2014. Cogswell, M., Lin, X., Purushwalkam, S., and Batra, D. Combining the best of graphical models and convnets for semantic segmentation. arXiv:1412.4313 , 2014. Dai, J., He, K., and Sun, J. Convolutional feature masking for joint object and stuff segmentation. arXiv:1412.1283 , 2014. Delong, A., Osokin, A., Isack, H. N., and Boykov, Y . Fast approximate energy minimization with label costs. IJCV , 2012. Eigen, D. and Fergus, R. Predicting depth, surface normals and semantic labels with a common

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

41

label costs. IJCV , 2012. Eigen, D. and Fergus, R. Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture. arXiv:1411.4734 , 2014. Everingham, M., Eslami, S. M. A., Gool, L. V ., Williams, C. K. I., Winn, J., and Zisserma, A. The pascal visual object classes challenge a retrospective. IJCV , 2014. Farabet, C., Couprie, C., Najman, L., and LeCun, Y . Learning hierarchical features for scene labeling.PAMI , 2013. Geiger, D. and Girosi, F. Parallel and deterministic algorithms from mrfs: Surface reconstruction. PAMI , 13(5):401–412, 1991. Geiger, D. and Yuille, A. A common framework for image segmentation. IJCV , 6(3):227–243, 1991. Girshick, R., Donahue, J., Darrell, T., and Malik, J. Rich feature hierarchies for accurate object detection and semantic segmentation. In CVPR , 2014.

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

42

detection and semantic segmentation. In CVPR , 2014. Giusti, A., Ciresan, D., Masci, J., Gambardella, L., and Schmidhuber, J. Fast image scanning with deep max-pooling convolutional neural networks. In ICIP , 2013. Gonfaus, J. M., Boix, X., Van de Weijer, J., Bagdanov, A. D., Serrat, J., and Gonzalez, J. Harmony potentials for joint classification and segmentation. In CVPR , 2010. Hariharan, B., Arbel ´aez, P., Bourdev, L., Maji, S., and Malik, J. Semantic contours from inverse detectors. In ICCV , 2011. Hariharan, B., Arbel ´aez, P., Girshick, R., and Malik, J. Hypercolumns for object segmentation and fine-grained localization. arXiv:1411.5752 , 2014a. Hariharan, B., Arbel ´aez, P., Girshick, R., and Malik, J. Simultaneous detection and segmentation. InECCV , 2014b. 12 Published as a conference paper at ICLR 2015

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

43

InECCV , 2014b. 12 Published as a conference paper at ICLR 2015 He, X., Zemel, R. S., and Carreira-Perpindn, M. Multiscale conditional random fields for image labeling. In CVPR , 2004. Jia, Y ., Shelhamer, E., Donahue, J., Karayev, S., Long, J., Girshick, R., Guadarrama, S., and Darrell, T. Caffe: Convolutional architecture for fast feature embedding. arXiv:1408.5093 , 2014. Kohli, P., Ladicky, L., and Torr, P. H. Robust higher order potentials for enforcing label consistency. IJCV , 2009. Kokkinos, I., Deriche, R., Faugeras, O., and Maragos, P. Computational analysis and learning for a biologically motivated model of boundary detection. Neurocomputing , 71(10):1798–1812, 2008. Kr¨ahenb ¨uhl, P. and Koltun, V . Efficient inference in fully connected crfs with gaussian edge potentials. In NIPS , 2011.

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

44

Kr¨ahenb ¨uhl, P. and Koltun, V . Parameter learning and convergent inference for dense random fields. InICML , 2013. Krizhevsky, A., Sutskever, I., and Hinton, G. E. Imagenet classification with deep convolutional neural networks. In NIPS , 2013. Ladicky, L., Russell, C., Kohli, P., and Torr, P. H. Associative hierarchical crfs for object class image segmentation. In ICCV , 2009. LeCun, Y ., Bottou, L., Bengio, Y ., and Haffner, P. Gradient-based learning applied to document recognition. In Proc. IEEE , 1998. Lempitsky, V ., Vedaldi, A., and Zisserman, A. Pylon model for semantic segmentation. In NIPS , 2011. Long, J., Shelhamer, E., and Darrell, T. Fully convolutional networks for semantic segmentation. arXiv:1411.4038 , 2014. Lucchi, A., Li, Y ., Boix, X., Smith, K., and Fua, P. Are spatial and global constraints really necessary

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

45

Lucchi, A., Li, Y ., Boix, X., Smith, K., and Fua, P. Are spatial and global constraints really necessary for segmentation? In ICCV , 2011. Mallat, S. A Wavelet Tour of Signal Processing . Acad. Press, 2 edition, 1999. Mostajabi, M., Yadollahpour, P., and Shakhnarovich, G. Feedforward semantic segmentation with zoom-out features. arXiv:1412.0774 , 2014. Papandreou, G., Kokkinos, I., and Savalle, P.-A. Untangling local and global deformations in deep convolutional networks for image classification and sliding window detection. arXiv:1412.0296 , 2014. Papandreou, G., Chen, L.-C., Murphy, K., and Yuille, A. L. Weakly- and semi-supervised learning of a DCNN for semantic image segmentation. arXiv:1502.02734 , 2015. Rother, C., Kolmogorov, V ., and Blake, A. Grabcut: Interactive foreground extraction using iterated graph cuts. In SIGGRAPH , 2004.

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

46

Rother, C., Kolmogorov, V ., and Blake, A. Grabcut: Interactive foreground extraction using iterated graph cuts. In SIGGRAPH , 2004. Sermanet, P., Eigen, D., Zhang, X., Mathieu, M., Fergus, R., and LeCun, Y . Overfeat: Integrated recognition, localization and detection using convolutional networks. arXiv:1312.6229 , 2013. Shotton, J., Winn, J., Rother, C., and Criminisi, A. Textonboost for image understanding: Multiclass object recognition and segmentation by jointly modeling texture, layout, and context. IJCV , 2009. Simonyan, K. and Zisserman, A. Very deep convolutional networks for large-scale image recognition. arXiv:1409.1556 , 2014. Szegedy, C., Liu, W., Jia, Y ., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V ., and Rabinovich, A. Going deeper with convolutions. arXiv:1409.4842 , 2014. 13 Published as a conference paper at ICLR 2015

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

47

Rabinovich, A. Going deeper with convolutions. arXiv:1409.4842 , 2014. 13 Published as a conference paper at ICLR 2015 Tompson, J., Jain, A., LeCun, Y ., and Bregler, C. Joint Training of a Convolutional Network and a Graphical Model for Human Pose Estimation. In NIPS , 2014. Uijlings, J., van de Sande, K., Gevers, T., and Smeulders, A. Selective search for object recognition. IJCV , 2013. Wang, P., Shen, X., Lin, Z., Cohen, S., Price, B., and Yuille, A. Towards unified depth and semantic prediction from a single image. In CVPR , 2015. Yadollahpour, P., Batra, D., and Shakhnarovich, G. Discriminative re-ranking of diverse segmentations. In CVPR , 2013. Zeiler, M. D. and Fergus, R. Visualizing and understanding convolutional networks. In ECCV , 2014. Zhang, N., Donahue, J., Girshick, R., and Darrell, T. Part-based r-cnns for fine-grained category detection. In ECCV , 2014.

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

48

detection. In ECCV , 2014. Zheng, S., Jayasumana, S., Romera-Paredes, B., Vineet, V ., Su, Z., Du, D., Huang, C., and Torr, P. Conditional random fields as recurrent neural networks. arXiv:1502.03240 , 2015. 14

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

1502.03167

0

arXiv:1502.03167v3 [cs.LG] 2 Mar 2015BatchNormalization: AcceleratingDeepNetworkTrainingb y ReducingInternalCovariateShift Sergey Ioffe GoogleInc., sioffe@google.comChristianSzegedy GoogleInc., szegedy@google.com Abstract TrainingDeepNeuralNetworksiscomplicatedbythefact that the distributionofeach layer’sinputschangesduring training, as the parametersof the previouslayers change. This slows down the training by requiringlower learning ratesandcarefulparameterinitialization,andmakesitno toriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Ourmethoddrawsitsstrengthfrommakingnormalizationapartofthemodelarchitectureandperformingthe normalization for each training mini-batch . Batch Normalizationallowsustousemuchhigherlearningratesand be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalizationachievesthe same accuracy with 14 times fewer training steps, and beats the original model

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

1

Batch Normalizationachievesthe same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batchnormalizednetworks,weimproveuponthebestpublished result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracyofhumanraters. 1 Introduction Deep learning has dramatically advanced the state of the art in vision, speech, and many other areas. Stochastic gradient descent (SGD) has proved to be an effective way of training deep networks, and SGD variants such as momentum (Sutskeveret al., 2013) and Adagrad (Duchiet al.,2011)havebeenusedtoachievestate ofthe art performance. SGD optimizes the parameters Θof the network,soasto minimizetheloss Θ = argmin Θ1 NN/summationdisplay i=1ℓ(xi,Θ)

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

2

network,soasto minimizetheloss Θ = argmin Θ1 NN/summationdisplay i=1ℓ(xi,Θ) wherex1...Nisthetrainingdataset. With SGD,thetrainingproceedsinsteps,andateachstepweconsidera minibatchx1...mofsizem. The mini-batchis usedtoapproximate the gradient of the loss functionwith respect to the parameters,bycomputing 1 m∂ℓ(xi,Θ) ∂Θ.Usingmini-batchesofexamples,asopposedtooneexampleatatime,ishelpfulinseveralways. First,thegradient ofthelossoveramini-batchisanestimateofthegradient overthetrainingset, whose qualityimprovesas thebatch size increases. Second, computation over a batch can be much more efficient than mcomputations for individual examples, due to the parallelism afforded by the modern computingplatforms. While stochastic gradient is simple and effective, it requires careful tuning of the model hyper-parameters, specificallythelearningrateusedinoptimization,aswell

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

3

requires careful tuning of the model hyper-parameters, specificallythelearningrateusedinoptimization,aswell as the initial values for the model parameters. The trainingiscomplicatedbythefactthattheinputstoeachlayer areaffectedbytheparametersofallprecedinglayers–so that small changes to the network parameters amplify as thenetworkbecomesdeeper. The change in the distributions of layers’ inputs presents a problem because the layers need to continuously adapt to the new distribution. When the input distributiontoalearningsystemchanges,itissaidtoexperiencecovariateshift (Shimodaira, 2000). This is typically handled via domain adaptation (Jiang, 2008). However, the notion of covariate shift can be extended beyond the learningsystemasawhole,toapplytoitsparts,suchasa sub-networkora layer. Considera networkcomputing ℓ=F2(F1(u,Θ1),Θ2) whereF1andF2are arbitrary transformations, and the parameters Θ1,Θ2are to be learned so as to minimize the lossℓ. Learning Θ2can be viewed as if the inputs x =F1(u,Θ1)arefedintothesub-network

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

4

the lossℓ. Learning Θ2can be viewed as if the inputs x =F1(u,Θ1)arefedintothesub-network ℓ=F2(x,Θ2). Forexample,agradientdescentstep Θ2←Θ2−α mm/summationdisplay i=1∂F2(xi,Θ2) ∂Θ2 (forbatchsize mandlearningrate α)isexactlyequivalent to that for a stand-alone network F2with input x. Therefore, the input distribution properties that make training more efficient – such as having the same distribution between the training and test data – apply to training the sub-network as well. As such it is advantageous for the distributionof xtoremainfixedovertime. Then, Θ2does 1 not have to readjust to compensate for the change in the distributionof x. Fixed distribution of inputs to a sub-network would havepositiveconsequencesforthelayers outsidethesubnetwork,as well. Consider a layer with a sigmoid activation function z =g(Wu+b)whereuis the layer input, the weight matrix Wand bias vector bare the layer parameters to be learned, and g(x) =1

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

5

the weight matrix Wand bias vector bare the layer parameters to be learned, and g(x) =1 1+exp(−x). As|x| increases, g′(x)tends to zero. This means that for all dimensionsof x =Wu+bexceptthosewithsmallabsolute values,thegradientflowingdownto uwillvanishandthe model will train slowly. However, since xis affected by W,band the parameters of all the layers below, changes tothoseparametersduringtrainingwilllikelymovemany dimensions of xinto the saturated regime of the nonlinearity and slow down the convergence. This effect is amplified as the network depth increases. In practice, the saturation problem and the resulting vanishing gradientsareusuallyaddressedbyusingRectifiedLinearUnits (Nair&Hinton, 2010) ReLU(x) = max( x,0), careful initialization (Bengio&Glorot, 2010; Saxeet al., 2013), and small learning rates. If, however, we could ensure that the distribution of nonlinearity inputs remains more stable as the network trains, then the optimizer would be less likely to get stuck in the saturated regime, and the trainingwouldaccelerate. We refer to the change in the distributions of internal

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

6

less likely to get stuck in the saturated regime, and the trainingwouldaccelerate. We refer to the change in the distributions of internal nodes of a deep network, in the course of training, as Internal Covariate Shift . Eliminating it offers a promise of faster training. We propose a new mechanism, which we callBatch Normalization , that takes a step towards reducing internal covariate shift, and in doing so dramatically accelerates the training of deep neural nets. It accomplishes this via a normalization step that fixes the meansandvariancesoflayerinputs. BatchNormalization also has a beneficial effect on the gradient flow through the network, by reducing the dependence of gradients on the scale of the parameters or of their initial values. This allows us to use much higher learning rates without the risk of divergence. Furthermore, batch normalization regularizes the model and reduces the need for Dropout(Srivastavaet al., 2014). Finally, Batch Normalization makes it possible to use saturating nonlinearities by preventingthe network from getting stuck in the saturatedmodes. In Sec. 4.2, we apply Batch Normalization to the bestperforming ImageNet classification network, and show that we can match its performance using only 7% of the

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

7

that we can match its performance using only 7% of the training steps, and can further exceed its accuracy by a substantial margin. Using an ensemble of such networks trained with Batch Normalization, we achieve the top-5 error rate that improves upon the best known results on ImageNetclassification.2 Towards Reducing Internal CovariateShift We define Internal Covariate Shift as the change in the distribution of network activations due to the change in networkparametersduringtraining. Toimprovethetraining, we seek to reduce the internal covariate shift. By fixingthe distributionof the layer inputs xas the training progresses,weexpecttoimprovethetrainingspeed. Ithas been long known (LeCunetal., 1998b; Wiesler &Ney, 2011) that the network training convergesfaster if its inputsarewhitened–i.e.,linearlytransformedtohavezero meansandunitvariances,anddecorrelated. Aseachlayer observestheinputsproducedbythelayersbelow,itwould be advantageousto achieve the same whiteningof the inputsof each layer. By whitening the inputsto each layer,

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

8

be advantageousto achieve the same whiteningof the inputsof each layer. By whitening the inputsto each layer, we would take a step towards achieving the fixed distributions of inputs that would remove the ill effects of the internalcovariateshift. We couldconsiderwhiteningactivationsat everytraining step or at some interval, either by modifying the network directly or by changing the parameters of the optimization algorithm to depend on the network activation values (Wiesleret al., 2014; Raikoetal., 2012; Poveyet al., 2014; Desjardins&Kavukcuoglu). However, if these modifications are interspersed with the optimization steps, then the gradient descent step may attempt to update the parameters in a way that requires the normalization to be updated, which reduces the effect of the gradient step. For example, consider a layer with the input uthat addsthe learned bias b, and normalizes the result by subtracting the mean of the activation computed over the training data: /hatwidex=x−E[x]where x=u+b,X={x1...N}is the set of values of xover the training set, and E [x] =1 N/summationtextN i=1xi. If a gradient

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

9

the training set, and E [x] =1 N/summationtextN i=1xi. If a gradient descent step ignores the dependence of E [x]onb, then it will update b←b+ ∆b, where∆b∝−∂ℓ/∂/hatwidex. Then u+ (b+ ∆b)−E[u+ (b+ ∆b)] =u+b−E[u+b]. Thus, the combination of the update to band subsequent change in normalization led to no change in the output of the layer nor, consequently, the loss. As the training continues, bwill grow indefinitely while the loss remains fixed. Thisproblemcangetworseifthenormalizationnot only centers but also scales the activations. We have observed this empirically in initial experiments, where the model blows up when the normalization parameters are computedoutsidethe gradientdescentstep. The issue with the above approach is that the gradient descent optimization does not take into account the fact that the normalization takes place. To address this issue, we would like to ensure that, for any parameter values, the network alwaysproducesactivationswith the desired distribution. Doing so would allow the gradient of the

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

10

we would like to ensure that, for any parameter values, the network alwaysproducesactivationswith the desired distribution. Doing so would allow the gradient of the loss with respect to the model parameters to account for the normalization, and for its dependence on the model parameters Θ. Let again xbe a layer input, treated as a 2 vector, andXbe the set of these inputs over the training dataset. Thenormalizationcanthenbewrittenasatransformation /hatwidex =Norm(x,X) which depends not only on the given training example x but on all examples X– each of which depends on Θif xis generatedby anotherlayer. For backpropagation,we wouldneedtocomputetheJacobians ∂Norm(x,X) ∂xand∂Norm(x,X) ∂X; ignoring the latter term would lead to the explosion describedabove. Withinthisframework,whiteningthelayer inputs is expensive, as it requires computing the covariance matrix Cov [x] =Ex∈X[xxT]−E[x]E[x]Tand its inverse square root, to produce the whitened activations Cov[x]−1/2(x−E[x]), as well as the derivatives of these

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

11

inverse square root, to produce the whitened activations Cov[x]−1/2(x−E[x]), as well as the derivatives of these transformsforbackpropagation.Thismotivatesustoseek an alternative that performs input normalization in a way that is differentiable and does not require the analysis of theentiretrainingset aftereveryparameterupdate. Some of the previous approaches (e.g. (Lyu&Simoncelli, 2008)) use statistics computed over a single training example, or, in the case of image networks, over differentfeature maps at a given location. However, this changes the representation ability of a network by discarding the absolute scale of activations. We want to a preservethe informationin the network,by normalizing the activations in a training example relative tothe statisticsoftheentiretrainingdata. 3 Normalization via Mini-Batch Statistics Since the full whitening of each layer’s inputs is costly and not everywhere differentiable, we make two necessary simplifications. The first is that instead of whitening the features in layer inputs and outputs jointly, we will normalizeeachscalarfeatureindependently,bymakingit have the mean of zero and the variance of 1. For a layer

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

12

normalizeeachscalarfeatureindependently,bymakingit have the mean of zero and the variance of 1. For a layer withd-dimensionalinput x = (x(1)...x(d)),wewillnormalizeeachdimension /hatwidex(k)=x(k)−E[x(k)]/radicalbig Var[x(k)] wheretheexpectationandvariancearecomputedoverthe trainingdataset. Asshownin(LeCunetal.,1998b),such normalizationspeedsupconvergence,evenwhenthefeaturesarenotdecorrelated. Notethatsimplynormalizingeachinputofalayermay change what the layer can represent. For instance, normalizing the inputsof a sigmoid wouldconstrain them to the linear regime of the nonlinearity. To address this, we makesurethat thetransformationinsertedin thenetwork can represent the identity transform . To accomplish this,weintroduce,foreachactivation x(k),apairofparameters γ(k),β(k),whichscale andshift thenormalizedvalue: y(k)=γ(k)/hatwidex(k)+β(k). These parameters are learned along with the original model parameters, and restore the representation power

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

13

y(k)=γ(k)/hatwidex(k)+β(k). These parameters are learned along with the original model parameters, and restore the representation power ofthenetwork. Indeed,bysetting γ(k)=/radicalbig Var[x(k)]and β(k)=E[x(k)], we couldrecoverthe originalactivations, ifthatwerethe optimalthingto do. Inthebatchsettingwhereeachtrainingstepisbasedon theentire trainingset, we woulduse the wholeset to normalize activations. However,this is impracticalwhen using stochastic optimization. Therefore, we make the secondsimplification: since we use mini-batchesin stochastic gradient training, each mini-batch produces estimates ofthemeanandvariance ofeachactivation. Thisway,the statistics used for normalization can fully participate in the gradient backpropagation. Note that the use of minibatchesis enabledbycomputationof per-dimensionvariances rather than joint covariances; in the joint case, regularizationwouldbe requiredsince the mini-batchsize is likely to be smaller than the number of activations being whitened,resultinginsingularcovariancematrices.

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

14

likely to be smaller than the number of activations being whitened,resultinginsingularcovariancematrices. Consider a mini-batch Bof sizem. Since the normalization is applied to each activation independently, let us focusonaparticularactivation x(k)andomitkforclarity. We havemvaluesofthisactivationinthemini-batch, B={x1...m}. Letthenormalizedvaluesbe /hatwidex1...m,andtheirlineartransformationsbe y1...m. We referto thetransform BNγ,β:x1...m→y1...m as theBatch Normalizing Transform . We present the BN TransforminAlgorithm1. Inthealgorithm, ǫisaconstant addedtothemini-batchvariancefornumericalstability. Input:Valuesof xovera mini-batch: B={x1...m}; Parametersto belearned: γ,β Output:{yi=BNγ,β(xi)} µB←1 mm/summationdisplay i=1xi // mini-batchmean σ2 B←1 mm/summationdisplay i=1(xi−µB)2// mini-batchvariance

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

15

σ2 B←1 mm/summationdisplay i=1(xi−µB)2// mini-batchvariance /hatwidexi←xi−µB/radicalbig σ2 B+ǫ// normalize yi←γ/hatwidexi+β≡BNγ,β(xi) // scale andshift Algorithm 1: Batch Normalizing Transform, applied to activation xoveramini-batch. TheBNtransformcanbeaddedtoanetworktomanipulate any activation. In the notation y=BNγ,β(x), we 3 indicate that the parameters γandβare to be learned, but it should be noted that the BN transform does not independently process the activation in each training example. Rather, BN γ,β(x)depends both on the training exampleand the other examples in the mini-batch . The scaled and shifted values yare passed to other network layers. The normalized activations /hatwidexare internal to our transformation, but their presence is crucial. The distributions of values of any /hatwidexhas the expected value of 0 and the variance of 1, as long as the elements of each mini-batch are sampled from the same distribution, and

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

16

and the variance of 1, as long as the elements of each mini-batch are sampled from the same distribution, and if we neglect ǫ. This can be seen by observing that/summationtextm i=1/hatwidexi= 0and1 m/summationtextm i=1/hatwidex2 i= 1, and taking expectations. Eachnormalizedactivation /hatwidex(k)canbeviewedas an input to a sub-network composed of the linear transformy(k)=γ(k)/hatwidex(k)+β(k), followed by the other processing doneby the originalnetwork. Thesesub-network inputs all have fixed means and variances, and although the jointdistributionofthese normalized /hatwidex(k)canchange over the course of training, we expect that the introduction of normalized inputs accelerates the training of the sub-networkand,consequently,thenetworkasawhole. During training we need to backpropagate the gradient of loss ℓthrough this transformation,as well as compute the gradients with respect to the parameters of the BN transform. We use chainrule,as follows(beforesimplification): ∂ℓ ∂/hatwidexi=∂ℓ ∂yi·γ

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

17

∂ℓ ∂/hatwidexi=∂ℓ ∂yi·γ ∂ℓ ∂σ2 B=/summationtextm i=1∂ℓ ∂/hatwidexi·(xi−µB)·−1 2(σ2 B+ǫ)−3/2 ∂ℓ ∂µB=/parenleftbigg/summationtextm i=1∂ℓ ∂/hatwidexi·−1√ σ2 B+ǫ/parenrightbigg +∂ℓ ∂σ2 B·/summationtextm i=1−2(xi−µB) m ∂ℓ ∂xi=∂ℓ ∂/hatwidexi·1√ σ2 B+ǫ+∂ℓ ∂σ2 B·2(xi−µB) m+∂ℓ ∂µB·1 m ∂ℓ ∂γ=/summationtextm i=1∂ℓ ∂yi·/hatwidexi ∂ℓ

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

18

∂ℓ ∂γ=/summationtextm i=1∂ℓ ∂yi·/hatwidexi ∂ℓ ∂β=/summationtextm i=1∂ℓ ∂yi Thus,BNtransformisadifferentiabletransformationthat introduces normalized activations into the network. This ensures that as the model is training, layers can continue learningoninputdistributionsthatexhibitlessinternal covariate shift, thus accelerating the training. Furthermor e, the learned affine transform applied to these normalized activationsallowstheBNtransformtorepresenttheidentity transformationandpreservesthenetworkcapacity. 3.1 Training and Inference with BatchNormalizedNetworks ToBatch-Normalize anetwork,wespecifyasubsetofactivations and insert the BN transform for each of them, according to Alg. 1. Any layer that previously received xas the input, now receives BN (x). A model employing Batch Normalization can be trained using batch gradient descent,orStochasticGradientDescentwithamini-batch sizem >1, or with any of its variants such as Adagrad(Duchiet al.,2011). Thenormalizationofactivationsthat

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

19

sizem >1, or with any of its variants such as Adagrad(Duchiet al.,2011). Thenormalizationofactivationsthat dependsonthemini-batchallowsefficienttraining,but is neithernecessarynordesirableduringinference;wewant the output to depend only on the input, deterministically. For this, once the network has been trained, we use the normalization /hatwidex=x−E[x]/radicalbig Var[x]+ǫ using the population, rather than mini-batch, statistics. Neglecting ǫ, these normalized activations have the same mean0 and variance1 as duringtraining. We use the unbiased variance estimate Var [x] =m m−1·EB[σ2 B], where theexpectationisovertrainingmini-batchesofsize mand σ2 Baretheirsamplevariances. Usingmovingaveragesinstead, we can track the accuracy of a model as it trains. Sincethemeansandvariancesarefixedduringinference, the normalization is simply a linear transform applied to eachactivation. Itmayfurtherbecomposedwiththescaling byγand shift by β, to yield a single linear transform

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

20

the normalization is simply a linear transform applied to eachactivation. Itmayfurtherbecomposedwiththescaling byγand shift by β, to yield a single linear transform that replacesBN (x). Algorithm 2 summarizesthe procedurefortrainingbatch-normalizednetworks. Input:NetworkNwith trainableparameters Θ; subsetofactivations {x(k)}K k=1 Output: Batch-normalizednetworkforinference, Ninf BN 1:Ntr BN←N// TrainingBN network 2:fork= 1...Kdo 3:Add transformation y(k)=BNγ(k),β(k)(x(k))to Ntr BN(Alg.1) 4:Modify each layer in Ntr BNwith input x(k)to take y(k)instead 5:end for 6:TrainNtr BNto optimize the parameters Θ∪ {γ(k),β(k)}K k=1 7:Ninf BN←Ntr BN// InferenceBN networkwithfrozen // parameters 8:fork= 1...Kdo 9:// Forclarity, x≡x(k),γ≡γ(k),µB≡µ(k) B, etc.

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

21

9:// Forclarity, x≡x(k),γ≡γ(k),µB≡µ(k) B, etc. 10:Process multiple training mini-batches B, each of sizem,andaverageoverthem: E[x]←EB[µB] Var[x]←m m−1EB[σ2 B] 11:InNinf BN, replace the transform y=BNγ,β(x)with y=γ√ Var[x]+ǫ·x+/parenleftbig β−γE[x]√ Var[x]+ǫ/parenrightbig 12:end for Algorithm2: Traininga Batch-NormalizedNetwork 3.2 Batch-Normalized Convolutional Networks Batch Normalization can be applied to any set of activations in the network. Here, we focus on transforms 4 that consist of an affine transformation followed by an element-wisenonlinearity: z =g(Wu+b) whereWandbare learned parametersof the model, and g(·)isthenonlinearitysuchassigmoidorReLU.Thisformulation covers both fully-connected and convolutional layers. We add the BN transform immediately before the

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

22

layers. We add the BN transform immediately before the nonlinearity,bynormalizing x =Wu+b. Wecouldhave also normalized the layer inputs u, but since uis likely the output of another nonlinearity, the shape of its distributionislikelytochangeduringtraining,andconstrainin g its first and second moments would not eliminate the covariate shift. In contrast, Wu + bis more likely to have a symmetric,non-sparsedistribution,that is “moreGaussian”(Hyv¨ arinen&Oja,2000);normalizingitislikelyto produceactivationswithastable distribution. Notethat,sincewenormalize Wu+b,thebiasbcanbe ignoredsinceitseffectwillbecanceledbythesubsequent meansubtraction(theroleofthebiasissubsumedby βin Alg.1). Thus, z =g(Wu+b)is replacedwith z =g(BN(Wu)) where the BN transformis applied independentlyto each dimension of x =Wu, with a separate pair of learned parameters γ(k),β(k)perdimension. Forconvolutionallayers,we additionallywant the normalization to obey the convolutional property – so that

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

23

parameters γ(k),β(k)perdimension. Forconvolutionallayers,we additionallywant the normalization to obey the convolutional property – so that different elements of the same feature map, at different locations, are normalized in the same way. To achieve this, we jointly normalize all the activations in a minibatch, overall locations. In Alg. 1, we let Bbe the set of all values in a feature map across both the elements of a mini-batch and spatial locations – so for a mini-batch of sizemand feature maps of size p×q, we use the effective mini-batch of size m′=|B|=m·pq. We learn a pair of parameters γ(k)andβ(k)per feature map, rather than per activation. Alg. 2 is modified similarly, so that duringinferencetheBNtransformappliesthesamelinear transformationtoeachactivationina givenfeaturemap. 3.3 Batch Normalization enables higher learning rates In traditional deep networks, too-high learning rate may result in the gradients that explode or vanish, as well as getting stuck in poor local minima. Batch Normalization helps address these issues. By normalizing activations throughout the network, it prevents small changes

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

24

getting stuck in poor local minima. Batch Normalization helps address these issues. By normalizing activations throughout the network, it prevents small changes to the parameters from amplifying into larger and suboptimal changes in activations in gradients; for instance, it prevents the training from getting stuck in the saturated regimesofnonlinearities. BatchNormalizationalsomakestrainingmoreresilient totheparameterscale. Normally,largelearningratesmay increasethescaleoflayerparameters,whichthenamplifythegradientduringbackpropagationandleadtothemodel explosion. However, with Batch Normalization, backpropagation through a layer is unaffected by the scale of itsparameters. Indeed,fora scalar a, BN(Wu) =BN((aW)u) andwe canshowthat ∂BN((aW)u) ∂u=∂BN(Wu) ∂u ∂BN((aW)u) ∂(aW)=1 a·∂BN(Wu) ∂W The scale does not affect the layer Jacobian nor, consequently, the gradient propagation. Moreover, larger weights lead to smallergradients, and Batch Normalizationwill stabilize theparametergrowth. We further conjecture that Batch Normalization may

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

25

weights lead to smallergradients, and Batch Normalizationwill stabilize theparametergrowth. We further conjecture that Batch Normalization may leadthelayerJacobianstohavesingularvaluescloseto1, which is known to be beneficial for training (Saxeet al., 2013). Consider two consecutive layers with normalized inputs, and the transformation between these normalized vectors:/hatwidez =F(/hatwidex). Ifweassumethat /hatwidexand/hatwidezareGaussian anduncorrelated,andthat F(/hatwidex)≈J/hatwidexisalineartransformationforthe givenmodelparameters,thenboth /hatwidexand/hatwidez have unit covariances, and I=Cov[/hatwidez] =JCov[/hatwidex]JT= JJT. Thus,JJT=I, and so all singular values of J are equal to 1, which preserves the gradient magnitudes during backpropagation. In reality, the transformation is notlinear,andthenormalizedvaluesarenotguaranteedto be Gaussian nor independent, but we nevertheless expect Batch Normalization to help make gradient propagation better behaved. The precise effect of Batch Normalization on gradient propagation remains an area of further study.

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

26

Batch Normalization to help make gradient propagation better behaved. The precise effect of Batch Normalization on gradient propagation remains an area of further study. 3.4 Batch Normalization regularizes the model When training with Batch Normalization, a training example is seen in conjunction with other examples in the mini-batch, and the training network no longer producing deterministic values for a given training example. In our experiments,we foundthis effect to be advantageous to the generalization of the network. Whereas Dropout (Srivastavaet al., 2014) is typically used to reduce overfitting,inabatch-normalizednetworkwefoundthatitcan beeitherremovedorreducedinstrength. 4 Experiments 4.1 Activationsovertime To verify the effects of internal covariate shift on training, and the ability of Batch Normalization to combat it, weconsideredtheproblemofpredictingthedigitclasson theMNISTdataset(LeCunetal.,1998a). Weusedavery simple network, with a 28x28binary image as input, and 5 10K20K30K40K50K0.70.80.91 Without BN With BN −202 −202 (a) (b)WithoutBN (c)With BN

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

27

Without BN With BN −202 −202 (a) (b)WithoutBN (c)With BN Figure 1: (a) The test accuracy of the MNIST network trained with and without Batch Normalization, vs. the number of training steps. Batch Normalization helps the network train faster and achieve higher accuracy. (b, c)The evolution of input distributions to a typical sigmoid,overthecourseoftraining,shownas {15,50,85}th percentiles. Batch Normalization makes the distribution morestableandreducestheinternalcovariateshift. 3fully-connectedhiddenlayerswith100activationseach. Eachhiddenlayercomputes y =g(Wu+b)withsigmoid nonlinearity, and the weights Winitialized to small random Gaussian values. The last hidden layer is followed by a fully-connected layer with 10 activations (one per class) and cross-entropyloss. We trained the network for 50000steps, with 60 examplespermini-batch. We added BatchNormalizationtoeachhiddenlayerofthenetwork, as in Sec. 3.1. We were interested in the comparison betweenthebaselineandbatch-normalizednetworks,rather thanachievingthestateoftheartperformanceonMNIST

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

28

thanachievingthestateoftheartperformanceonMNIST (whichthe describedarchitecturedoesnot). Figure 1(a) shows the fraction of correct predictions by the two networks on held-out test data, as training progresses. The batch-normalized network enjoys the higher test accuracy. To investigate why, we studied inputs to the sigmoid, in the original network Nand batchnormalizednetwork Ntr BN(Alg.2)overthecourseoftraining. InFig.1(b,c)weshow,foronetypicalactivationfrom the last hidden layer of each network, how its distribution evolves. The distributions in the original network change significantly over time, both in their mean and the variance, which complicates the training of the subsequent layers. In contrast, the distributions in the batch normalizednetworkaremuchmorestableastrainingprogresses,whichaidsthe training. 4.2 ImageNetclassification We applied Batch Normalization to a new variant of the Inception network (Szegedyetal., 2014), trained on the ImageNet classification task (Russakovskyet al., 2014). The network has a large number of convolutional and pooling layers, with a softmax layer to predict the image class, out of 1000 possibilities. Convolutional layers use

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

29

The network has a large number of convolutional and pooling layers, with a softmax layer to predict the image class, out of 1000 possibilities. Convolutional layers use ReLU asthenonlinearity. Themaindifferenceto thenetwork described in (Szegedyet al., 2014) is that the 5×5 convolutionallayersare replacedby two consecutivelayers of3×3convolutionswith up to 128filters. The network contains 13.6·106parameters, and, other than the top softmax layer, has no fully-connected layers. Moredetails are given in the Appendix. We refer to this model asInception intherestofthetext. Themodelwastrained using a version of Stochastic Gradient Descent with momentum(Sutskeveretal.,2013),usingthemini-batchsize of32. Thetrainingwasperformedusingalarge-scale,distributed architecture (similar to (Deanet al., 2012)). All networksare evaluatedastrainingprogressesbycomputing the validation accuracy @1, i.e. the probability of predicting the correct label out of 1000 possibilities, on aheld-outset,usinga singlecropperimage. Inourexperiments,weevaluatedseveralmodifications

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

30

aheld-outset,usinga singlecropperimage. Inourexperiments,weevaluatedseveralmodifications ofInceptionwithBatchNormalization. Inallcases,Batch Normalizationwasappliedtotheinputofeachnonlinearity, in a convolutional way, as described in section 3.2, whilekeepingtherestofthearchitectureconstant. 4.2.1 AcceleratingBN Networks SimplyaddingBatchNormalizationtoanetworkdoesnot take full advantage of our method. To do so, we further changed the network and its training parameters, as follows: Increase learning rate. In a batch-normalized model, we have been able to achieve a training speedup from higherlearningrates,with noill sideeffects(Sec.3.3). RemoveDropout. As describedin Sec. 3.4, Batch NormalizationfulfillssomeofthesamegoalsasDropout. Removing Dropout from Modified BN-Inception speeds up training,withoutincreasingoverfitting. Reduce the L2weight regularization. While in Inception anL2loss on the model parameters controls overfitting, in Modified BN-Inception the weight of this loss is

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

31

reduced by a factor of 5. We find that this improves the accuracyontheheld-outvalidationdata. Accelerate the learning rate decay. In training Inception, learning rate was decayed exponentially. Because our network trains faster than Inception, we lower the learningrate 6timesfaster. Remove Local Response Normalization While Inception and other networks (Srivastavaet al., 2014) benefit from it, we found that with Batch Normalization it is not necessary. Shuffletrainingexamplesmorethoroughly. Weenabled within-shardshufflingofthetrainingdata,whichprevents thesameexamplesfromalwaysappearinginamini-batch together. This led to about 1% improvements in the validation accuracy, which is consistent with the view of Batch Normalization as a regularizer (Sec. 3.4): the randomization inherent in our method should be most beneficialwhenitaffectsanexampledifferentlyeachtimeitis seen. Reduce the photometric distortions. Because batchnormalized networks train faster and observe each trainingexamplefewertimes,weletthetrainerfocusonmore “real”imagesbydistortingthemless. 6

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

32

“real”imagesbydistortingthemless. 6 5M 10M 15M 20M 25M 30M0.40.50.60.70.8 Inception BN−Baseline BN−x5 BN−x30 BN−x5−Sigmoid Steps to match Inception Figure 2: Single crop validation accuracy of Inception and its batch-normalized variants, vs. the number of trainingsteps.Model Stepsto72.2% Maxaccuracy Inception 31.0·10672.2% BN-Baseline 13.3·10672.7% BN-x5 2.1·10673.0% BN-x30 2.7·10674.8% BN-x5-Sigmoid 69.8% Figure 3: For Inception and the batch-normalized variants, the number of training steps required to reach the maximum accuracy of Inception(72.2%), and the maximum accuracy achieved by the network. 4.2.2 Single-NetworkClassification We evaluated the following networks, all trained on the LSVRC2012 training data, and tested on the validation data: Inception : the network described at the beginning of Section4.2,trainedwiththeinitiallearningrateof0.001 5.

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

33

data: Inception : the network described at the beginning of Section4.2,trainedwiththeinitiallearningrateof0.001 5. BN-Baseline : Same as Inception with Batch Normalizationbeforeeachnonlinearity. BN-x5: Inception with Batch Normalization and the modifications in Sec. 4.2.1. The initial learning rate was increased by a factor of 5, to 0.0075. The same learning rateincreasewithoriginalInceptioncausedthemodelparameterstoreachmachineinfinity. BN-x30: LikeBN-x5, but with the initial learning rate 0.045(30timesthatofInception). BN-x5-Sigmoid : LikeBN-x5, but with sigmoid nonlinearityg(t) =1 1+exp(−x)instead of ReLU. We also attempted to train the original Inception with sigmoid, but themodelremainedat theaccuracyequivalenttochance. In Figure 2, we show the validation accuracy of the networks, as a function of the number of training steps. Inception reached the accuracy of 72.2% after 31·106 training steps. The Figure 3 shows, for each network, the number of training steps required to reach the same

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

34

Inception reached the accuracy of 72.2% after 31·106 training steps. The Figure 3 shows, for each network, the number of training steps required to reach the same 72.2%accuracy,aswellasthemaximumvalidationaccuracy reached by the network and the number of steps to reachit. By onlyusingBatch Normalization( BN-Baseline ),we matchtheaccuracyofInceptioninlessthanhalfthenumber of training steps. By applying the modifications in Sec. 4.2.1, we significantly increase the training speed of the network. BN-x5needs 14 times fewer steps than Inception to reach the 72.2% accuracy. Interestingly, increasing the learning rate further ( BN-x30) causes the model to train somewhat slowerinitially, but allows it to reachahigherfinalaccuracy. Itreaches74.8%after 6·106 steps, i.e. 5 times fewer steps than required by Inception toreach72.2%. We also verified that the reduction in internal covariate shift allows deep networks with Batch Normalizationto be trained when sigmoid is used as the nonlinearity,

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

35

We also verified that the reduction in internal covariate shift allows deep networks with Batch Normalizationto be trained when sigmoid is used as the nonlinearity, despite the well-known difficulty of training such networks. Indeed, BN-x5-Sigmoid achieves the accuracy of 69.8%. WithoutBatchNormalization,Inceptionwithsigmoidneverachievesbetterthan 1/1000accuracy. 4.2.3 Ensemble Classification The current reported best results on the ImageNet Large Scale Visual RecognitionCompetitionare reachedby the Deep Image ensemble of traditional models (Wuet al., 2015) and the ensemble model of (Heet al., 2015). The latterreportsthetop-5errorof4.94%,asevaluatedbythe ILSVRCserver. Herewereportatop-5validationerrorof 4.9%, and test error of 4.82% (according to the ILSVRC server). This improvesupon the previousbest result, and exceedstheestimatedaccuracyofhumanratersaccording to(Russakovskyet al.,2014). Forourensemble,weused6networks. Eachwasbased onBN-x30,modifiedviasomeofthefollowing: increased

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

36

Forourensemble,weused6networks. Eachwasbased onBN-x30,modifiedviasomeofthefollowing: increased initial weights in the convolutionallayers; using Dropout (with the Dropout probability of 5% or 10%, vs. 40% for the original Inception); and using non-convolutional, per-activation Batch Normalization with last hidden layers of the model. Each network achieved its maximum accuracyafter about 6·106training steps. The ensemble prediction was based on the arithmetic average of class probabilities predicted by the constituent networks. The detailsofensembleandmulticropinferencearesimilarto (Szegedyet al., 2014). We demonstrate in Fig. 4 that batch normalization allowsusto set new state-of-the-artby a healthymarginon theImageNetclassificationchallengebenchmarks. 5 Conclusion We have presented a novel mechanism for dramatically accelerating the training of deep networks. It is based on the premise that covariate shift, which is known to complicate the trainingof machine learning systems, also ap7 Model Resolution Crops Models Top-1error Top-5error GoogLeNetensemble 224 144 7 - 6.67%

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

37

Model Resolution Crops Models Top-1error Top-5error GoogLeNetensemble 224 144 7 - 6.67% DeepImagelow-res 256 - 1 - 7.96% DeepImagehigh-res 512 - 1 24.88 7.42% DeepImageensemble variable - - - 5.98% BN-Inceptionsinglecrop 224 1 1 25.2% 7.82% BN-Inceptionmulticrop 224 144 1 21.99% 5.82% BN-Inceptionensemble 224 144 6 20.1% 4.9%* Figure 4: Batch-Normalized Inception comparison with previous stat e of the art on the provided validation set comprising50000images. *BN-Inceptionensemblehasreached4 .82%top-5erroronthe100000imagesofthetestsetof theImageNetasreportedbythe test server. plies to sub-networks and layers, and removing it from internal activations of the network may aid in training. Our proposed method draws its power from normalizing activations, and from incorporating this normalization in the network architecture itself. This ensures that the normalization is appropriately handled by any optimization method that is being used to train the network. To enable stochastic optimization methods commonly used in deep network training, we perform the normalization for

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

38

method that is being used to train the network. To enable stochastic optimization methods commonly used in deep network training, we perform the normalization for eachmini-batch,andbackpropagatethegradientsthrough the normalization parameters. Batch Normalization adds only two extra parameters per activation, and in doing so preserves the representation ability of the network. We presentedanalgorithmforconstructing,training,andper forming inference with batch-normalized networks. The resulting networks can be trained with saturating nonlinearities, are more tolerant to increased training rates, an d oftendonotrequireDropoutforregularization. Merely adding Batch Normalization to a state-of-theartimageclassificationmodelyieldsasubstantialspeedup in training. By further increasing the learning rates, removing Dropout, and applying other modifications afforded by Batch Normalization, we reach the previous state of the art with onlya small fractionof trainingsteps –andthenbeatthestateoftheartinsingle-networkimage classification. Furthermore, by combining multiple models trained with Batch Normalization, we perform better thanthebestknownsystemonImageNet,byasignificant margin.

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

39

thanthebestknownsystemonImageNet,byasignificant margin. Interestingly, our method bears similarity to the standardization layer of (G¨ ulc ¸ehre& Bengio, 2013), though the two methodsstem from very differentgoals, and perform different tasks. The goal of Batch Normalization is to achieve a stable distribution of activation values throughout training, and in our experiments we apply it before the nonlinearity since that is where matching the first and second moments is more likely to result in a stable distribution. On the contrary,(G¨ ulc ¸ehre&Bengio , 2013) apply the standardizationlayer to the outputof the nonlinearity, which results in sparser activations. In our large-scaleimage classification experiments,we havenot observedthenonlinearity inputstobesparse,neitherwith nor without Batch Normalization. Other notable differ-entiating characteristics of Batch Normalization include the learned scale and shift that allow the BN transform to representidentity (the standardizationlayer did not re quirethissinceitwasfollowedbythelearnedlineartransform that, conceptually, absorbs the necessary scale and

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

40

to representidentity (the standardizationlayer did not re quirethissinceitwasfollowedbythelearnedlineartransform that, conceptually, absorbs the necessary scale and shift), handling of convolutional layers, deterministic i nferencethatdoesnotdependonthemini-batch,andbatchnormalizingeachconvolutionallayerin thenetwork. In this work, we have not explored the full range of possibilitiesthatBatchNormalizationpotentiallyenabl es. Our future work includes applications of our method to Recurrent Neural Networks (Pascanuet al., 2013), where theinternalcovariateshiftandthevanishingorexploding gradients may be especially severe, and which would allowustomorethoroughlytestthehypothesisthatnormalizationimprovesgradientpropagation(Sec.3.3). Weplan toinvestigatewhetherBatchNormalizationcanhelpwith domain adaptation, in its traditional sense – i.e. whether the normalization performed by the network would allow it to more easily generalize to new data distributions,perhapswithjustarecomputationofthepopulation meansandvariances(Alg.2). Finally,webelievethatfurthertheoreticalanalysisofthealgorithmwouldallowstil l moreimprovementsandapplications. References

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

41

moreimprovementsandapplications. References Bengio, Yoshua and Glorot, Xavier. Understanding the difficultyoftrainingdeepfeedforwardneuralnetworks. InProceedings of AISTATS 2010 , volume 9, pp. 249– 256,May2010. Dean,Jeffrey,Corrado,GregS.,Monga,Rajat,Chen,Kai, Devin,Matthieu,Le,QuocV., Mao,MarkZ.,Ranzato, Marc’Aurelio,Senior,Andrew,Tucker,Paul,Yang,Ke, and Ng, Andrew Y. Large scale distributed deep networks. In NIPS,2012. Desjardins, Guillaume and Kavukcuoglu,Koray. Natural neuralnetworks. (unpublished). Duchi, John, Hazan, Elad, and Singer, Yoram. Adaptive subgradientmethodsfor onlinelearning and stochastic 8 optimization. J.Mach.Learn.Res. ,12:2121–2159,July 2011. ISSN1532-4435. G¨ ulc ¸ehre, C ¸aglar and Bengio, Yoshua. Knowledge matters: Importanceof prior informationfor optimization.

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

42

G¨ ulc ¸ehre, C ¸aglar and Bengio, Yoshua. Knowledge matters: Importanceof prior informationfor optimization. CoRR,abs/1301.4083,2013. He, K., Zhang, X., Ren, S., and Sun, J. Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification. ArXiv e-prints , February 2015. Hyv¨ arinen, A. and Oja, E. Independentcomponent analysis: Algorithms and applications. Neural Netw. , 13 (4-5):411–430,May2000. Jiang, Jing. A literature survey on domain adaptation of statistical classifiers, 2008. LeCun, Y., Bottou, L., Bengio, Y., and Haffner, P. Gradient-based learning applied to document recognition.Proceedings of the IEEE , 86(11):2278–2324, November1998a. LeCun, Y., Bottou, L., Orr, G., and Muller, K. Efficient backprop. InOrr,G.andK.,Muller(eds.), NeuralNetworks: Tricks ofthetrade .Springer,1998b.

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

43

backprop. InOrr,G.andK.,Muller(eds.), NeuralNetworks: Tricks ofthetrade .Springer,1998b. Lyu, S and Simoncelli, E P. Nonlinear image representation using divisive normalization. In Proc. Computer Vision and Pattern Recognition , pp. 1–8. IEEE Computer Society, Jun 23-28 2008. doi: 10.1109/CVPR. 2008.4587821. Nair,VinodandHinton,GeoffreyE. Rectifiedlinearunits improve restricted boltzmann machines. In ICML, pp. 807–814.Omnipress,2010. Pascanu, Razvan, Mikolov, Tomas, and Bengio, Yoshua. On the difficulty of training recurrent neural networks. InProceedingsofthe30thInternationalConferenceon MachineLearning,ICML 2013,Atlanta,GA,USA,1621June2013 ,pp.1310–1318,2013. Povey, Daniel, Zhang, Xiaohui, and Khudanpur, Sanjeev. Parallel training of deep neural networks with natural gradient and parameter averaging. CoRR, abs/1410.7455,2014.

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

44

natural gradient and parameter averaging. CoRR, abs/1410.7455,2014. Raiko, Tapani, Valpola, Harri, and LeCun, Yann. Deep learning made easier by linear transformations in perceptrons. In International Conference on Artificial IntelligenceandStatistics(AISTATS) ,pp.924–932,2012. Russakovsky,Olga,Deng,Jia,Su,Hao,Krause,Jonathan, Satheesh, Sanjeev, Ma, Sean, Huang, Zhiheng, Karpathy, Andrej, Khosla, Aditya, Bernstein, Michael, Berg, Alexander C., and Fei-Fei, Li. ImageNet Large Scale Visual RecognitionChallenge,2014.Saxe, Andrew M., McClelland, James L., and Ganguli, Surya. Exact solutions to the nonlinear dynamics of learning in deep linear neural networks. CoRR, abs/1312.6120,2013. Shimodaira, Hidetoshi. Improving predictive inference under covariate shift by weighting the log-likelihood function. JournalofStatisticalPlanningandInference , 90(2):227–244,October2000.

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

45

under covariate shift by weighting the log-likelihood function. JournalofStatisticalPlanningandInference , 90(2):227–244,October2000. Srivastava, Nitish, Hinton, Geoffrey, Krizhevsky, Alex, Sutskever, Ilya, and Salakhutdinov, Ruslan. Dropout: A simple way to preventneural networksfrom overfitting.J. Mach. Learn. Res. , 15(1):1929–1958, January 2014. Sutskever, Ilya, Martens, James, Dahl, George E., and Hinton, Geoffrey E. On the importance of initialization and momentum in deep learning. In ICML (3), volume 28 of JMLR Proceedings , pp. 1139–1147. JMLR.org,2013. Szegedy, Christian, Liu, Wei, Jia, Yangqing, Sermanet, Pierre, Reed, Scott, Anguelov, Dragomir, Erhan, Dumitru, Vanhoucke, Vincent, and Rabinovich, Andrew. Going deeper with convolutions. CoRR, abs/1409.4842,2014.

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

46

abs/1409.4842,2014. Wiesler, Simon and Ney, Hermann. A convergenceanalysis of log-lineartraining. In Shawe-Taylor,J., Zemel, R.S.,Bartlett,P.,Pereira,F.C.N.,andWeinberger,K.Q. (eds.),AdvancesinNeuralInformationProcessingSystems24,pp.657–665,Granada,Spain,December2011. Wiesler, Simon, Richard, Alexander, Schl¨ uter, Ralf, and Ney, Hermann. Mean-normalized stochastic gradient for large-scale deep learning. In IEEE International Conference on Acoustics, Speech, and Signal Processing,pp.180–184,Florence,Italy,May2014. Wu, Ren, Yan, Shengen, Shan, Yi, Dang, Qingqing, and Sun,Gang. Deepimage: Scalingupimagerecognition, 2015. Appendix Variantofthe Inception Model Used Figure 5 documents the changes that were performed compared to the architecture with respect to the GoogleNet archictecture. For the interpretation of this table, please consult (Szegedyetal., 2014). The notable architecture changes compared to the GoogLeNet model include:

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

47

table, please consult (Szegedyetal., 2014). The notable architecture changes compared to the GoogLeNet model include: •The 5×5 convolutional layers are replaced by two consecutive 3×3 convolutional layers. This increases the maximum depth of the network by 9 9 weight layers. Also it increases the number of parameters by 25% and the computational cost is increasedbyabout30%. •The number 28×28 inception modules is increased from2to 3. •Inside the modules, sometimes average, sometimes maximum-poolingis employed. This is indicated in theentriescorrespondingtothepoolinglayersofthe table. •There are no across the board pooling layers between any two Inception modules, but stride-2 convolution/pooling layers are employed before the filterconcatenationin themodules3c,4e. Our model employed separable convolution with depth multiplier 8on the first convolutionallayer. This reduces thecomputationalcost while increasingthememoryconsumptionat trainingtime. 10 typepatch size/ strideoutput sizedepth #1×1#3×3 reduce#3×3double#3×3 reducedouble #3×3Pool+proj

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

48

strideoutput sizedepth #1×1#3×3 reduce#3×3double#3×3 reducedouble #3×3Pool+proj convolution* 7×7/2112×112×641 maxpool 3×3/256×56×64 0 convolution 3×3/156×56×192 1 64 192 maxpool 3×3/228×28×192 0 inception (3a) 28×28×256 364 64 64 64 96 avg+ 32 inception (3b) 28×28×320 364 64 96 64 96 avg+ 64 inception (3c) stride 2 28×28×576 3 0128 160 64 96max +pass through inception (4a) 14×14×576 3224 64 96 96 128 avg+ 128 inception (4b) 14×14×576 3192 96 128 96 128 avg+ 128 inception (4c) 14×14×576 3160 128 160 128 160 avg+ 128 inception (4d) 14×14×576 396 128 192 160 192 avg+ 128 inception (4e) stride 2 14×14×1024 3 0128 192 192 256max +pass through inception (5a) 7×7×1024 3352 192 320 160 224 avg+ 128

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1502.03167

49

inception (5a) 7×7×1024 3352 192 320 160 224 avg+ 128 inception (5b) 7×7×1024 3352 192 320 192 224 max+ 128 avgpool 7×7/11×1×1024 0 Figure5: Inceptionarchitecture 11

1502.03167

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization. It also acts as a regularizer, in some cases eliminating the need for Dropout. Applied to a state-of-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.9% top-5 validation error (and 4.8% test error), exceeding the accuracy of human raters.

http://arxiv.org/pdf/1502.03167

cs.LG

20150211

20150302

1503.00848

0

1 Multiscale Combinatorial Grouping for Image Segmentation and Object Proposal Generation Jordi Pont-Tuset*, Pablo Arbel ´aez*, Jonathan T. Barron, Member, IEEE, Ferran Marques, Senior Member, IEEE, Jitendra Malik, Fellow, IEEE Abstract —We propose a unified approach for bottom-up hierarchical image segmentation and object proposal generation for recognition, called Multiscale Combinatorial Grouping (MCG). For this purpose, we first develop a fast normalized cuts algorithm. We then propose a high-performance hierarchical segmenter that makes effective use of multiscale information. Finally, we propose a grouping strategy that combines our multiscale regions into highly-accurate object proposals by exploring efficiently their combinatorial space. We also present Single-scale Combinatorial Grouping (SCG), a faster version of MCG that produces competitive proposals in under five second per image. We conduct an extensive and comprehensive empirical validation on the BSDS500, SegVOC12, SBD, and COCO datasets, showing that MCG produces state-of-the-art contours, hierarchical regions, and object proposals. Index Terms —Image segmentation, object proposals, normalized cuts. F 1 I NTRODUCTION

1503.00848

Multiscale Combinatorial Grouping for Image Segmentation and Object Proposal Generation

We propose a unified approach for bottom-up hierarchical image segmentation and object proposal generation for recognition, called Multiscale Combinatorial Grouping (MCG). For this purpose, we first develop a fast normalized cuts algorithm. We then propose a high-performance hierarchical segmenter that makes effective use of multiscale information. Finally, we propose a grouping strategy that combines our multiscale regions into highly-accurate object proposals by exploring efficiently their combinatorial space. We also present Single-scale Combinatorial Grouping (SCG), a faster version of MCG that produces competitive proposals in under five second per image. We conduct an extensive and comprehensive empirical validation on the BSDS500, SegVOC12, SBD, and COCO datasets, showing that MCG produces state-of-the-art contours, hierarchical regions, and object proposals.

http://arxiv.org/pdf/1503.00848

cs.CV

20150303

20160301

1503.00848

1

Index Terms —Image segmentation, object proposals, normalized cuts. F 1 I NTRODUCTION TWOparadigms have shaped the field of object recognition in the last decade. The first one, popularized by the Viola-Jones face detection algorithm [1], formulates object localization as window classification. The basic scanning-window architecture, relying on histograms of gradients and linear support vector machines, was introduced by Dalal and Triggs [2] in the context of pedestrian detection and is still at the core of seminal object detectors on the PASCAL challenge such as Deformable Part Models [3]. The second paradigm relies on perceptual grouping to provide a limited number of high-quality and categoryindependent object proposals, which can then be described with richer representations and used as input to more sophisticated learning methods. Examples in this family are [4], [5]. Recently, this approach has dominated the PASCAL segmentation challenge [6], [7], [8], [9], improved object detection [10], fine-grained categorization [11] and proven competitive in large-scale classification [12]. Since the power of this second paradigm is critically dependent on the accuracy and the number of object

1503.00848

Multiscale Combinatorial Grouping for Image Segmentation and Object Proposal Generation

We propose a unified approach for bottom-up hierarchical image segmentation and object proposal generation for recognition, called Multiscale Combinatorial Grouping (MCG). For this purpose, we first develop a fast normalized cuts algorithm. We then propose a high-performance hierarchical segmenter that makes effective use of multiscale information. Finally, we propose a grouping strategy that combines our multiscale regions into highly-accurate object proposals by exploring efficiently their combinatorial space. We also present Single-scale Combinatorial Grouping (SCG), a faster version of MCG that produces competitive proposals in under five second per image. We conduct an extensive and comprehensive empirical validation on the BSDS500, SegVOC12, SBD, and COCO datasets, showing that MCG produces state-of-the-art contours, hierarchical regions, and object proposals.

http://arxiv.org/pdf/1503.00848

cs.CV

20150303

20160301

1503.00848

2

[11] and proven competitive in large-scale classification [12]. Since the power of this second paradigm is critically dependent on the accuracy and the number of object proposals, an increasing body of research has delved J. Pont-Tuset and F. Marques are with the Department of Signal Theory and Communications, Universitat Polit` ecnica de Catalunya, BarcelonaTech (UPC), Spain. E-mail: fjordi.pont,ferran.marques g@upc.edu P . Arbel´ aez is with the Department of Biomedical Engineering, Universidad de los Andes, Colombia. E-mail: pa.arbelaez@uniandes.edu.co J. T. Barron, and J. Malik are with the Department of Electrical Engineering and Computer Science, University of California at Berkeley, Berkeley, CA 94720. E-mail: fbarron,malikg@eecs.berkeley.edu * The first two authors contributed equally Fig. 1. Top: original image, instance-level ground truth from COCO and our multiscale hierarchical segmentation. Bottom: our best object proposals among 150. into the problem of their generation [13], [14], [15], [12],

1503.00848

Multiscale Combinatorial Grouping for Image Segmentation and Object Proposal Generation

We propose a unified approach for bottom-up hierarchical image segmentation and object proposal generation for recognition, called Multiscale Combinatorial Grouping (MCG). For this purpose, we first develop a fast normalized cuts algorithm. We then propose a high-performance hierarchical segmenter that makes effective use of multiscale information. Finally, we propose a grouping strategy that combines our multiscale regions into highly-accurate object proposals by exploring efficiently their combinatorial space. We also present Single-scale Combinatorial Grouping (SCG), a faster version of MCG that produces competitive proposals in under five second per image. We conduct an extensive and comprehensive empirical validation on the BSDS500, SegVOC12, SBD, and COCO datasets, showing that MCG produces state-of-the-art contours, hierarchical regions, and object proposals.

http://arxiv.org/pdf/1503.00848

cs.CV

20150303

20160301

1503.00848

3

into the problem of their generation [13], [14], [15], [12], [16], [17], [18], [19]. However, those approaches typically focus on learning generic properties of objects from a set of examples, while reasoning on a fixed set of regions and contours produced by external bottom-up segmenters such as [20], [21]. In this paper, we propose a unified approach to multiscale hierarchical segmentation and object proposal generation called Multiscale Combinatorial Grouping (MCG). Fig. 1 shows an example of our results and Fig. 2 an overview of our pipeline. Our main contributions are: An efficient normalized cuts algorithm, which in practice provides a 20speed-up to the eigenvector computation required for contour globalization [20], [22] (Sect. 3.1). A state-of-the-art hierarchical segmenter that leverages multiscale information (Sect. 3.3). A grouping algorithm that produces accurate object proposals by efficiently exploring the combinatorial space of our multiscale regions (Sect. 5).arXiv:1503.00848v4 [cs.CV] 1 Mar 2016 2

1503.00848

Multiscale Combinatorial Grouping for Image Segmentation and Object Proposal Generation

We propose a unified approach for bottom-up hierarchical image segmentation and object proposal generation for recognition, called Multiscale Combinatorial Grouping (MCG). For this purpose, we first develop a fast normalized cuts algorithm. We then propose a high-performance hierarchical segmenter that makes effective use of multiscale information. Finally, we propose a grouping strategy that combines our multiscale regions into highly-accurate object proposals by exploring efficiently their combinatorial space. We also present Single-scale Combinatorial Grouping (SCG), a faster version of MCG that produces competitive proposals in under five second per image. We conduct an extensive and comprehensive empirical validation on the BSDS500, SegVOC12, SBD, and COCO datasets, showing that MCG produces state-of-the-art contours, hierarchical regions, and object proposals.

http://arxiv.org/pdf/1503.00848

cs.CV

20150303

20160301

1503.00848

4

space of our multiscale regions (Sect. 5).arXiv:1503.00848v4 [cs.CV] 1 Mar 2016 2 F ixed-Scale Segmentation R escaling & A lignment C ombination R es oluti on C ombinatorial Grouping I mage P yramid Segmentati on P yramid Al i gned H iera rchies Object Proposals Multiscale Hierarchy Fig. 2. Multiscale Combinatorial Grouping . Starting from a multiresolution image pyramid, we perform hierarchical segmentation at each scale independently. We align these multiple hierarchies and combine them into a single multiscale segmentation hierarchy. Our grouping component then produces a ranked list of object proposals by efficiently exploring the combinatorial space of these regions. We conduct a comprehensive and large-scale empirical validation. On the BSDS500 (Sect. 4) we report significant progress in contour detection and hierarchical segmentation. On the VOC2012, SBD, and COCO segmentation datasets (Sect. 6), our proposals obtain overall state-ofthe-art accuracy both as segmented proposals and as

1503.00848

Multiscale Combinatorial Grouping for Image Segmentation and Object Proposal Generation

We propose a unified approach for bottom-up hierarchical image segmentation and object proposal generation for recognition, called Multiscale Combinatorial Grouping (MCG). For this purpose, we first develop a fast normalized cuts algorithm. We then propose a high-performance hierarchical segmenter that makes effective use of multiscale information. Finally, we propose a grouping strategy that combines our multiscale regions into highly-accurate object proposals by exploring efficiently their combinatorial space. We also present Single-scale Combinatorial Grouping (SCG), a faster version of MCG that produces competitive proposals in under five second per image. We conduct an extensive and comprehensive empirical validation on the BSDS500, SegVOC12, SBD, and COCO datasets, showing that MCG produces state-of-the-art contours, hierarchical regions, and object proposals.

http://arxiv.org/pdf/1503.00848

cs.CV

20150303

20160301

1503.00848

5

datasets (Sect. 6), our proposals obtain overall state-ofthe-art accuracy both as segmented proposals and as bounding boxes. MCG is efficient, its good generalization power makes it parameter free in practice, and it provides a ranked set of proposals that are competitive in all regimes of number of proposals. 2 R ELATED WORK For space reasons, we focus our review on recent normalized cut algorithms and object proposals for recognition. Fast normalized cuts : The efficient computation of normalized-cuts eigenvectors has been the subject of recent work, as it is often the computational bottleneck in grouping algorithms. Taylor [23] presented a technique for using a simple watershed oversegmentation to reduce the size of the eigenvector problem, sacrificing accuracy for speed. We take a similar approach of solving the eigenvector problem in a reduced space, though we use simple image-pyramid operations on the affinity matrix (instead of a separate segmentation algorithm) and we see no loss in performance despite a 20speed improvement. Maire and Yu [24] presented a novel multigrid solver for producing eigenvectors at multiple scales, which speeds up fine-scale eigenvector computation by leveraging coarse-scale solutions. Our

1503.00848

Multiscale Combinatorial Grouping for Image Segmentation and Object Proposal Generation

We propose a unified approach for bottom-up hierarchical image segmentation and object proposal generation for recognition, called Multiscale Combinatorial Grouping (MCG). For this purpose, we first develop a fast normalized cuts algorithm. We then propose a high-performance hierarchical segmenter that makes effective use of multiscale information. Finally, we propose a grouping strategy that combines our multiscale regions into highly-accurate object proposals by exploring efficiently their combinatorial space. We also present Single-scale Combinatorial Grouping (SCG), a faster version of MCG that produces competitive proposals in under five second per image. We conduct an extensive and comprehensive empirical validation on the BSDS500, SegVOC12, SBD, and COCO datasets, showing that MCG produces state-of-the-art contours, hierarchical regions, and object proposals.

http://arxiv.org/pdf/1503.00848

cs.CV

20150303

20160301