Datasets:

jamescalam

/

llama-2-arxiv-papers-chunked

like 17

1503.00848

6

multiple scales, which speeds up fine-scale eigenvector computation by leveraging coarse-scale solutions. Our technique also uses the scale-space structure of an image, but instead of solving the problem at multiple scales, we simply reduce the scale of the problem, solve it at a reduced scale, and then upsample the solution while preserving the structure of the image. As such, our technique is faster and much simpler, requiring only a few lines of code wrapped around a standard sparse eigensolver.Object Proposals : Class-independent methods that generate object hypotheses can be divided into those whose output is an image window and those that generate segmented proposals. Among the former, Alexe et al. [16] propose an objectness measure to score randomly-sampled image windows based on low-level features computed on the superpixels of [21]. Manen et al. [25] propose to use the Randomized Prim’s algorithm, Zitnick et al. [26] group contours directly to produce object windows, and Cheng et al. [27] generate box proposals at 300 images per second. In contrast to these approaches, we focus on the finer-grained task of pixel-accurate object extraction,

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

7

second. In contrast to these approaches, we focus on the finer-grained task of pixel-accurate object extraction, rather than on window selection. However, by just taking the bounding box around our segmented proposals, our results are also state of the art as window proposals. Among the methods that produce segmented proposals, Carreira and Sminchisescu [18] hypothesize a set of placements of fore- and background seeds and, for each configuration, solve a constrained parametric min-cut (CPMC) problem to generate a pool of object hypotheses. Endres and Hoiem [19] base their category-independent object proposals on an iterative generation of a hierarchy of regions, based on the contour detector of [20] and occlusion boundaries of [28]. Kim and Grauman [17] propose to match parts of the shape of exemplar objects, regardless of their class, to detected contours by [20]. They infer the presence and shape of a proposal object by adapting the matched object to the computed superpixels. Uijlings et al. [12] present a selective search algorithm based on segmentation. Starting with the superpixels of [21] for a variety of color spaces, they produce a set of segmentation hierarchies by region merging, which are

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

8

based on segmentation. Starting with the superpixels of [21] for a variety of color spaces, they produce a set of segmentation hierarchies by region merging, which are used to produce a set of object proposals. While we also take advantage of different hierarchies to gain diversity, we leverage multiscale information rather than different 3 color spaces. Recently, two works proposed to train a cascade of classifiers to learn which sets of regions should be merged to form objects. Ren and Shankhnarovich [29] produce full region hierarchies by iteratively merging pairs of regions and adapting the classifiers to different scales. Weiss and Taskar [30] specialize the classifiers also to size and class of the annotated instances to produce object proposals. Malisiewicz and Efros [4] took one of the first steps towards combinatorial grouping, by running multiple segmenters with different parameters and merging up to three adjacent regions. In [8], another step was taken by considering hierarchical segmentations at three different scales and combining pairs and triplets of adjacent regions from the two coarser scales to produce object proposals. The most recent wave of object proposal algorithms

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

9

regions from the two coarser scales to produce object proposals. The most recent wave of object proposal algorithms is represented by [13], [14], and [15], which all keep the quality of the seminal proposal works while improving the speed considerably. Kr ¨ahenb ¨uhl and Koltun [13] find object proposal by identifying critical level sets in geodesic distance transforms, based on seeds placed in learnt places in the image. Rantalankila et al. [14] perform a global and local search in the space of sets of superpixels. Humayun et al. [15] reuse a graph to perform many parametric min-cuts over different seeds in order to speed the process up. A substantial difference between our approach and previous work is that, instead of relying on precomputed hierarchies or superpixels, we propose a unified approach that produces and groups high-quality multiscale regions. With respect to the combinatorial approaches of [4], [8], our main contribution is to develop efficient algorithms to explore a much larger combinatorial space by taking into account a set of object examples, increasing thus the likelihood of having complete objects in the pool of proposals. Our approach has therefore

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

10

increasing thus the likelihood of having complete objects in the pool of proposals. Our approach has therefore the flexibility to adapt to specific applications and types of objects, and can produce proposals at any trade-off between their number and their accuracy. 3 T HESEGMENTATION ALGORITHM Consider a segmentation of the image into regions that partition its domain S=fSigi. A segmentation hierarchy is a family of partitions fS;S1;:::;SLgsuch that: (1) Sis the finest set of superpixels , (2)SLis the complete domain, and (3) regions from coarse levels are unions of regions from fine levels. A hierarchy where each level Si is assigned a real-valued index ican be represented by a dendrogram, a region tree where the height of each node is its index. Furthermore, it can also be represented as an ultrametric contour map (UCM), an image obtained by weighting the boundary of each pair of adjacent regions in the hierarchy by the index at which they are merged [31], [32]. This representation unifies the problems of contour detection and hierarchical image segmentation: * ## #

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

11

[31], [32]. This representation unifies the problems of contour detection and hierarchical image segmentation: * ## # Fig. 3. Duality between a UCM and a region tree : Schematic view of the dual representation of a segmentation hierarchy as a region dendrogram and as an ultrametric contour map. a threshold at level iin the UCM produces the segmentationSi. Figure 3 schematizes these concepts. First, the lower left corner shows the probability of boundary of a UCM. One of the main properties of a UCM is that when we threshold the contour strength at a certain value, we obtain a closed boundary map, and thus a partition. Thresholding at different i, therefore, we obtain the so-called merging-sequence partitions (left column in Figure 3); named after the fact that a step in this sequence corresponds to merging the set of regions sharing the boundary of strength exactly i. For instance, the boundary between the wheels and the floor has strength 1, thus thresholding the contour above1makes the wheels merge with the floor. If we represent the regions in a partition as nodes of a graph, we can then represent the result of merging them as their

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

12

represent the regions in a partition as nodes of a graph, we can then represent the result of merging them as their parent in a tree. The result of sweeping all ivalues can therefore be represented as a region tree, whose root is the region representing the whole image (right part of Figure 3). Given that each merging is associated with a contour strength, the region tree is in fact a region dendogram. As an example, in the gPb-ucm algorithm of [20], brightness, color and texture gradients at three fixed disk sizes are first computed. These local contour cues are globalized using spectral graph-partitioning, resulting in the gPb contour detector. Hierarchical segmentation is then performed by iteratively merging adjacent regions based on the average gPb strength on their common boundary. This algorithm produces therefore a tree of regions at multiple levels of homogeneity in brightness, color and texture, and the boundary strength of its UCM can be interpreted as a measure of contrast. Coarse-to-fine is a powerful processing strategy in computer vision. We exploit it in two different ways to develop an efficient, scalable and high-performance segmentation algorithm: (1) To speed-up spectral graph

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

13

to develop an efficient, scalable and high-performance segmentation algorithm: (1) To speed-up spectral graph partitioning and (2) To create aligned segmentation hierarchies. 4 3.1 Fast Downsampled Eigenvector Computation The normalized cuts criterion is a key globalization mechanism of recent high-performance contour detectors such as [20], [22]. Although powerful, such spectral graph partitioning has a significant computational cost and memory footprint that limit its scalability. In this section, we present an efficient normalized cuts approximation which in practice preserves full performance for contour detection, while having low memory requirements and providing a 20speed-up. Given a symmetric affinity matrix A, we would like to compute the ksmallest eigenvectors of the Laplacian ofA. Directly computing such eigenvectors can be very costly even with sophisticated solvers, due to the large size ofA. We therefore present a technique for efficiently approximating the eigenvector computation by taking advantage of the multiscale nature of our problem: A models affinities between pixels in an image, and images naturally lend themselves to multiscale or pyramid-like

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

14

models affinities between pixels in an image, and images naturally lend themselves to multiscale or pyramid-like representations and algorithms. Our algorithm is inspired by two observations: 1) if Ais bistochastic (the rows and columns of Asum to 1) then the eigenvectors of the Laplacian Aare equal to the eigenvectors of the Laplacian of A2, and 2) because of the scale-similar nature of images, the eigenvectors of a “downsampled” version of Ain which every other pixel has been removed should be similar to the eigenvectors of A. Let us define pixel decimate (A), which takes an affinity matrix Aand returns the set of indices of rows/columns in Acorresponding to a decimated version of the image from which Awas constructed. That is, ifi=pixel decimate (A), thenA[i;i]is a decimated matrix in which alternating rows and columns of the imagehave been removed. Computing the eigenvectors of A[i;i]works poorly, as decimation disconnects pixels in the affinity matrix, but the eigenvectors of the decimated

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

15

A[i;i]works poorly, as decimation disconnects pixels in the affinity matrix, but the eigenvectors of the decimated squared affinity matrix A2[i;i]are similar to those of A, because by squaring the matrix before decimation we intuitively allow each pixel to propagate information to all of its neighbors in the graph, maintaining connections even after decimation. Our algorithm works by efficiently computing A2[i;i]asA[:;i]TA[:;i]1(the naive approach of first squaring Aand then decimating it is prohibitively expensive), computing the eigenvectors of A2[i;i], and then “upsampling” those eigenvectors back to the space of the original image by pre-multiplying byA[:;i]. This squaring-and-decimation procedure can be applied recursively several times, each application improving efficiency while slightly sacrificing accuracy. Pseudocode for our algorithm, which we call “DNCuts” (Downsampled Normalized Cuts) is given in

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

16

Pseudocode for our algorithm, which we call “DNCuts” (Downsampled Normalized Cuts) is given in Algorithm 1, where Ais our affinity matrix and dis the number of times that our squaring-and-decimation operation is applied. Our algorithm repeatedly applies our joint squaring-and-decimation procedure, computes 1. The Matlab-like notation A[:; i]indicates keeping the columns of matrix Awhose indices are in the set i.Algorithm 1 dncuts (A;d;k ) 1:A0 A 2:fors= [1;2;:::;d ]do 3:is pixel decimate (As1) 4:Bs As1[ :; is] 5:Cs diag (Bs~1)1Bs 6:As CT sBs 7:end for 8:Xd ncuts (Ad;k) 9:fors= [d;d1;:::; 1]do 10:Xs1 CsXs 11:end for 12:return whiten (X0) Fig. 4. Example of segmentation projection . In order to “snap” the boundaries of a segmentation R(left) to those

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

17

Fig. 4. Example of segmentation projection . In order to “snap” the boundaries of a segmentation R(left) to those of a segmentation S(middle), since they do not align, we compute(R;S)(right) by assigning to each segment in Sits mode among the labels of R. the smallest keigenvectors of the final “downsampled” matrix Adby using a standard sparse eigensolver ncuts (Ad;k), and repeatedly “upsamples” those eigenvectors. Because our Ais not bistochastic and decimation is not an orthonormal operation, we must do some normalization throughout the algorithm (line 5) and whiten the resulting eigenvectors (line 10). We found that values ofd= 2 ord= 3 worked well in practice. Larger values ofdyielded little speed improvement (as much of the cost is spent downsampling A0) and start negatively affecting accuracy. Our technique is similar to Nystrom’s method for computing the eigenvectors of a subset of A, but our squaring-and-decimation procedure means that we do not depend on long-range connections between pixels. 3.2 Aligning Segmentation Hierarchies

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

18

but our squaring-and-decimation procedure means that we do not depend on long-range connections between pixels. 3.2 Aligning Segmentation Hierarchies In order to leverage multi-scale information, our approach combines segmentation hierarchies computed independently at multiple image resolutions. However, since subsampling an image removes details and smooths away boundaries, the resulting UCMs are misaligned, as illustrated in the second panel of Fig. 2. In this section, we propose an algorithm to align an arbitrary segmentation hierarchy to a target segmentation and, in Sect. 5, we show its effectiveness for multi-scale segmentation. The basic operation is to “snap” the boundaries of a segmentationR=fRigito a segmentation S=fSjgj, as illustrated in Fig. 4. For this purpose, we define L(Sj), 5 the new label of a region Sj2S, as the majority label of its pixels inR: L(Sj) = arg max ijSj\Rij jSjj(1) We call the segmentation defined by this new labeling of all the regions of Stheprojection ofRontoSand denote it by(R;S).

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

19

We call the segmentation defined by this new labeling of all the regions of Stheprojection ofRontoSand denote it by(R;S). In order to project an UCM onto a target segmentation S, which we denote (UCM;S), we project in turn each of the levels of the hierarchy onto S. Note that, since all the levels are projected onto the same segmentation, the family of projections is by construction a hierarchy of segmentations. This procedure is summarized in pseudo-code in Algorithm 2. Algorithm 2 UCM Rescaling and Alignment Require: An UCM with a set of levels [t1;:::;tK] Require: A target segmentation S 1:UCM 0 2:fort= [t1;:::;tK]do 3:S sampleHierarchy (UCM;t) 4:S rescaleSegmentation (S;S) 5:S (S;S) 6:contours extractBoundary (S) 7: UCM max(UCM;tcontours ) 8:end for 9:return UCM Observe that the routines sampleHierarchy and extractBoundary can be computed efficiently because

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

20

8:end for 9:return UCM Observe that the routines sampleHierarchy and extractBoundary can be computed efficiently because they involve only thresholding operations and connected components labeling. The complexity is thus dominated by rescaleSegmentation in Step 4, a nearest neighbor interpolation, and the projection in Step 5, which are computedKtimes. 3.3 Multiscale Hierarchical Segmentation Single-scale segmentation : We consider as input the following local contour cues: (1) brightness, color and texture differences in half-disks of three sizes [33], (2) sparse coding on patches [22], and (3) structured forest contours [34]. We globalize the contour cues independently using our fast eigenvector gradients of Sect. 3.1, combine global and local cues linearly, and construct an UCM based on the mean contour strength. We tried learning weights using gradient ascent on the F-measure on the training set [20], but evaluating the final hierarchies rather than open contours. We observed that this objective favors the quality of contours at the expense of regions and obtained better overall results by optimizing the Segmentation Covering metric [20].

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

21

objective favors the quality of contours at the expense of regions and obtained better overall results by optimizing the Segmentation Covering metric [20]. Hierarchy Alignment : We construct a multiresolution pyramid with Nscales by subsampling / supersampling the original image and applying our singlescale segmenter. In order to preserve thin structures and details, we declare as set of possible boundarylocations the finest superpixels in the highest-resolution. Then, applying recursively Algorithm 2, we project each coarser UCM onto the next finer scale until aligning it to the highest resolution superpixels. Multiscale Hierarchy : After alignment, we have a fixed set of boundary locations, and Nstrengths for each of them, coming from the different scales. We formulate this problem as binary boundary classification and train a classifier that combines these Nfeatures into a single probability of boundary estimation. We experimented with several learning strategies for combining UCM strengths: (a) Uniform weights transformed into probabilities with Platt’s method. (b) SVMs and logistic regression, with both linear and additive kernels. (c)

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

22

probabilities with Platt’s method. (b) SVMs and logistic regression, with both linear and additive kernels. (c) Random Forests. (d) The same algorithm as for singlescale. We found the results with all learning methods surprisingly similar, in agreement with the observation reported by [33]. This particular learning problem, with only a handful of dimensions and millions of data points, is relatively easy and performance is mainly driven by our already high-performing and well calibrated features. We therefore use the simplest option (a). 4 E XPERIMENTS ON THE BSDS500 We conduct extensive experiments on the BSDS500 [35], using the standard evaluation metrics and following the best practice rules of that dataset. We also report results with a recent evaluation metric Fop[36], [37], PrecisionRecall for objects and parts, using the publicly-available code. Single-scale Segmentation : Table 1-top shows the performance of our single-scale segmenter for different types of input contours on the validation set of the BSDS500. We obtain high-quality hierarchies for all the cues considered, showing the generality of our approach. Furthermore, when using them jointly (row ’Comb.’ in top panel), our segmenter outperforms the versions with individual cues, suggesting its ability to leverage

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

23

Furthermore, when using them jointly (row ’Comb.’ in top panel), our segmenter outperforms the versions with individual cues, suggesting its ability to leverage diversified inputs. In terms of efficiency, our fast normalized cuts algorithm provides an average 20 speedup over [20], starting from the same local cues, with no significant loss in accuracy and with a low memory footprint. Multiscale Segmentation : Table 1-bottom evaluates our full approach in the same experimental conditions as the upper panel. We observe a consistent improvement in performance in all the metrics for all the inputs, which validates our architecture for multiscale segmentation. We experimented with the range of scales and found N=f0:5;1;2gadequate for our purposes. A finer sampling or a wider range of scales did not provide noticeable improvements. We tested also two degraded versions of our system (not shown in the table). For the first one, we resized contours to the original image resolution, created UCMs and combined them with the same method as our final system. For the second one, we transformed per-scale UCMs to the original resolution, 6 Boundary Region

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

24

same method as our final system. For the second one, we transformed per-scale UCMs to the original resolution, 6 Boundary Region Fb Fop SC PRI VI Input ODS OIS ODS OIS ODS OIS ODS OIS ODS OISSingle-ScalePb [33] 0.702 0.733 0.334 0.370 0.577 0.636 0.801 0.847 1.692 1.490 SC [22] 0.697 0.725 0.264 0.306 0.540 0.607 0.777 0.835 1.824 1.659 SF [34] 0.719 0.737 0.338 0.399 0.582 0.651 0.803 0.851 1.608 1.432 Comb. 0.719 0.750 0.358 0.403 0.602 0.655 0.809 0.855 1.619 1.405MultiscalePb [33] 0.713 0.745 0.350 0.389 0.598 0.656 0.807 0.856 1.601 1.418 SC [22] 0.705 0.734 0.331 0.384 0.579 0.647 0.799 0.851 1.637 1.460

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

25

SC [22] 0.705 0.734 0.331 0.384 0.579 0.647 0.799 0.851 1.637 1.460 SF [34] 0.725 0.744 0.370 0.420 0.600 0.660 0.810 0.854 1.557 1.390 Comb. 0.725 0.757 0.371 0.408 0.611 0.670 0.813 0.862 1.548 1.367 TABLE 1 BSDS500 val set. Control experiments for single-scale (top) and multiscale (bottom) hierarchical segmentation with different input contour detectors but omitted the strength transfer to the finest superpixels before combining them. The first ablated version produces interpolation artifacts and smooths away details, while the second one suffers from misalignment. Both fail to improve performance over the single-scale result, which provides additional empirical support for our multiscale approach. We also observed a small degradation in performance when forcing the input contour detector to use only the original image resolution, which indicates the advantages of considering multiscale information at all stages of processing. Since there are no drastic changes in our results when taking as input the different individual cues or their combination, in the sequel we use the version with structured

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

26

Since there are no drastic changes in our results when taking as input the different individual cues or their combination, in the sequel we use the version with structured forests for efficiency reasons, which we denote MCGUCM-Our . Comparison with state-of-the-art. : Figure 5 compares our multiscale hierarchical segmenter MCG ( ) and our single-scale hierarchical segmenter SCG ( ) on the BSDS500 test set against all the methods for which there is publicly available code. We also compare to the recent ISCRA [29] hierarchies ( ), provided precomputed by the authors. We obtain consistently the best results to date on BSDS500 for all operating regimes, both in terms of boundary and region quality. Note that the range of object scales in the BSDS500 is limited, which translates into modest absolute gains from MCG ( ) with respect to SCG ( ) in terms of boundary evaluation (left-hand plot), but more significant improvements in terms of objects and parts (right-hand plot). We will also observe more substantial improvements with respect to gPb-UCM ( ) when we move to PASCAL, SBD, and COCO in Section 6 (e.g. see Fig. 9).

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

27

move to PASCAL, SBD, and COCO in Section 6 (e.g. see Fig. 9). Ground-Truth Hierarchy : In order to gain further insights, we transfer the strength of each ground-truth segmentation to our highest-resolution superpixels SN and construct a combined hierarchy. This approximation to the semantic hierarchy, Ground-Truth Hierarchy (GTH) in Fig. 5, is an upper-bound for our approach as both share the same boundary locations and the only difference is their strength. Since the strength of GTHis proportional to the number of subjects marking it, it provides naturally the correct semantic ordering, where outer object boundaries are stronger than internal parts. Recently, Maire et al. [38] developed an annotation tool where the user encodes explicitly the “perceptual strength” of each contour. Our approach provides an alternative where the semantic hierarchy is reconstructed by sampling flat annotations from multiple subjects. 5 O BJECT PROPOSAL GENERATION The image segmentation algorithm presented in the previous sections builds on low-level features, so its regions are unlikely to represent accurately complete objects with heterogeneous parts. In this context, object proposal techniques create a set of hypotheses, possibly overlapping, which are more likely to represent full

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

28

objects with heterogeneous parts. In this context, object proposal techniques create a set of hypotheses, possibly overlapping, which are more likely to represent full object instances. Our approach to object proposal generation is to combinatorially look for sets of regions from our segmentation hierarchies that merged together are likely to represent complete objects. In this section, we first describe the efficient computation of certain region descriptors on a segmentation tree. Then, we describe how we use these techniques to efficiently explore the sets of merged regions from the hierarchy. Finally, we explain how we train the parameters of our algorithm for object proposals and how we rank the candidates by their probability of representing an object. Fast computation of descriptors : Let us assume, for instance, we want to compute the area of all regions in the hierarchy. Intuitively, working strictly on the merging-sequence partitions, we would need to scan all pixels in all partitions. On the other hand, working on the region tree allows us to scan the image only once to compute the area of the leaves, and then propagate the area to all parents as the addition of the areas of their children. As a drawback, the algorithms become intricate in terms of coding and necessary data structures. Take, for

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

29

area to all parents as the addition of the areas of their children. As a drawback, the algorithms become intricate in terms of coding and necessary data structures. Take, for instance, the computation of the neighbors of a certain region, which is trivial via scanning the partition on the merging sequence (look for region labels in the adjacent boundary pixels), but need tailored data structures and algorithms in the region tree. Formally, let us assume the image has ppixels, and we build a hierarchy based on ssuperpixels (leaves of the tree), andmmergings (different UCM strength values). The cost of computing the area on all regions using the merging-sequence partitions will be the cost of scanning all pixels in these partitions, thus p
(m+1). In contrast, the cost using the region tree will involve scanning the image once, and then propagating the area values, so p+m, which is notably faster. We built tailored algorithms and data structures to compute the bounding box, perimeter, and neighbors of a region using the region tree representation. 7 Boundaries 0 0:10:20:30:40:50:60:70:80:9 100:10:20:30:40:50:60:70:80:91 RecallPrecision

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

30

RecallPrecision Human [0.81-0.21] ISCRA [0.72] GTH [0.96] NCuts [0.63] MCG-Our [0.74] EGB [0.61] SCG-Our [0.74] MShift [0.60] gPb-UCM [0.73] Quadtree [0.41]Objects and Parts 0 0:10:20:30:40:50:60:70:80:9 100:10:20:30:40:50:60:70:80:91 RecallPrecisionHuman [0.56-0.05] GTH [0.64] MCG-Our [0.38] SCG-Our [0.35] ISCRA [0.35] gPb-UCM [0.35] MShift [0.23] NCuts [0.21] EGB [0.16] Quadtree [0.06] Fig. 5. BSDS500 test set. Precision-Recall curves for boundaries [35] (left) and for objects and parts [36] (right). The marker on each curve is placed on the Optimal Dataset Scale (ODS), and its F measure is presented in brackets in the

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

31

marker on each curve is placed on the Optimal Dataset Scale (ODS), and its F measure is presented in brackets in the legend. The isolated red asterisks refer to the human performance assessed on the same image (one human partition against the rest of human annotations on the same image) and on a different image (one human partition against the human partitions of a different, randomly selected, image). Combinatorial Grouping of Proposals : We can cast object segmentation as selecting some regions in the hierarchy, or in other words, as a combinatorial optimization problem on the hierarchy. To illustrate this principle, Figure 6(a) shows the simplified representation of the hierarchy in Figure 3. Figure 6(b) and (c) show two object proposals, and their representation in the region hierarchy. R4 R5 R1R2R3R6 R9 R7 R1R2R3R4R8 R5R6 (a) R9 R7 R1R2R3R4R8 R5R6 (b) R9 R7 R1R2R3R4R8 R5R6 (c)

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

32

R5R6 (b) R9 R7 R1R2R3R4R8 R5R6 (c) Fig. 6. Object segmentation as combinatorial optimization : Examples of objects (b), (c), formed by selecting regions from a hierarchy (a). Since hierarchies are built taking only low-level features into account, and do not use semantic information, objects will usually not be optimally represented using a single region in the hierarchy. As an example, Figure 6(c) shows the optimum representation of the car, consisting of three regions. A sensible approach to create object proposals is therefore to explore the set of n-tuples of regions. The main idea behind MCG is to explore this set efficiently, taking advantage of the region tree representation, via the fastcomputation of region neighbors. The whole set of tuples, however, is huge, and so it is not feasible to explore it exhaustively. Our approach ranks the proposals using the height of their regions in the tree (UCM strength) and explores the tree from the top, but only down to a certain threshold. To speed the process up, we design a top-down algorithm to compute the region neighbors and thus only compute them down to a certain depth in the tree.

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

33

process up, we design a top-down algorithm to compute the region neighbors and thus only compute them down to a certain depth in the tree. To further improve the results, we not only consider then-tuples from the resulting MCG-UCM-Our hierarchy, but also the rest of hierarchies computed at different scales. As we will show in the experiments, diversity significantly improves the proposal results. Parameter Learning via Pareto Front Optimization : MCG takes a set of diverse hierarchies and computes the n-tuples up to a certain UCM strength. We can interpret then-tuples from each hierarchy as a ranked list of Ni proposals that are put together to create the final set of Npproposals. At training time, we would like to find, for different values of Np, the number of proposals from each ranked list Nisuch that the joint pool of Npproposals has the best achievable quality. We frame this learning problem as a Pareto front optimization [39], [40] with two conflicting objective functions: number of proposals and achievable quality. At test time, we select a working point on the Pareto front, represented by the Ni

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

34

and achievable quality. At test time, we select a working point on the Pareto front, represented by the Ni values, based either on the number of proposals Npwe can handle or on the minimum achievable quality our 8 application needs, and we combine the Nitop proposals from each hierarchy list. Formally, assuming Rranked lists Li, an exhaustive learning algorithm would consider all possible values of theR-tuplefN1;:::;NRg, whereNi2f0;:::;jLijg; adding up toQR 1jLijparameterizations to try, which is intractable in our setting. Figure 7 illustrates the learning process. To reduce the dimensionality of the search space, we start by selecting two ranked lists L1,L2(green curves) and we sample the list atSlevels of number of proposals (green dots). We then scan the full S2different parameterizations to combine the proposals from both (blue dots). In other words, we analyze the sets of proposals created by combining the top N1fromL1(green dots) and the top N2fromL2. 1011021031040.30.60.9 Number of proposalsQuality 1011021031040.30.60.9 Number of proposalsQuality

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

35

N2fromL2. 1011021031040.30.60.9 Number of proposalsQuality 1011021031040.30.60.9 Number of proposalsQuality 1011021031040.30.60.9 Number of proposalsQualityRanked lists of object proposals L 1 L RPareto front reduction of parameters {N 1 · · · N R} Fig. 7. Pareto front learning : Training the combinatorial generation of proposals using the Pareto front The key step of the optimization consists in discarding those parameterizations whose quality point is not in the Pareto front (red curve). (i.e., those parameterizations that can be substituted by another with better quality with the same number of proposals, or by one with the same quality with less proposals.) We sample the Pareto front toSpoints and we iterate the process until all the ranked lists are combined. Each point in the final Pareto front corresponds to a particular parameterization fN1;:::;NRg. At train time, we choose a point on this curve, either at a given number of proposals Ncor at the achievable quality we are interested in (black triangle) and store the parameters N1;:::;NR . At test time, we combine the

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

36

we are interested in (black triangle) and store the parameters N1;:::;NR . At test time, we combine the N1;:::;NR top proposals from each ranked list. The number of sampled configurations using the proposed algorithm is (R1)S2, that is, we have reduced an exponential problem ( SR) to a quadratic one. Regressed Ranking of Proposals : To further reduce the number of proposals, we train a regressor from low-level features, as in [18]. Since the proposals are all formed by a set of regions from a reduced set of hierarchies, we focus on features that can be computed efficiently in a bottom-up fashion, as explained previously. We compute the following features: Size and location : Area and perimeter of the candidate; area, position, and aspect ratio of the bounding box; and the area balance between the regions in thecandidate. Shape : Perimeter (and sum of contour strength) divided by the squared root of the area; and area of the region divided by that of the bounding box. Contours : Sum of contour strength at the boundaries, mean contour strength at the boundaries; minimum and maximum UCM threshold of appearance and

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

37

Contours : Sum of contour strength at the boundaries, mean contour strength at the boundaries; minimum and maximum UCM threshold of appearance and disappearance of the regions forming the candidate. We train a Random Forest using these features to regress the object overlap with the ground truth, and diversify the ranking based on Maximum Marginal Relevance measures [18]. We tune the random forest learning on half training set and validating on the other half. For the final results, we train on the training set and evaluate our proposals on the validation set of PASCAL 2012. 6 E XPERIMENTS ON PASCAL VOC, SBD, AND COCO This section presents our large-scale empirical validation of the object proposal algorithm described in the previous section. We perform experiments in three annotated databases, with a variety of measures that demonstrate the state-of-the-art performance of our algorithm. Datasets and Evaluation Measures : We conduct experiments in the following three annotated datasets: the segmentation challenge of PASCAL 2012 Visual Object Classes (SegVOC12) [41], the Berkeley Semantic Boundaries Dataset (SBD) [42], and the Microsoft Common Objects in Context (COCO) [43]. They all consist of images with annotated objects of different categories.

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

38

Objects in Context (COCO) [43]. They all consist of images with annotated objects of different categories. Table 2 summarizes the number of images and object instances in each database. Number of Number of Number of Classes Images Objects SegVOC12 20 2 913 9 847 SBD 20 12 031 32 172 COCO 80 123 287 910 983 TABLE 2 Sizes of the databases Regarding the performance metrics, we measure the achievable quality with respect to the number of proposals, that is, the quality we would have if an oracle selected the best proposal among the pool. This aligns with the fact that object proposals are a preprocessing step for other algorithms that will represent and classify them. We want, therefore, the achievable quality within the proposals to be as high as possible, while reducing the number of proposals to make the final system as fast as possible. As a measure of quality of a specific proposal with respect to an annotated object, we consider the Jaccard indexJ, also known as overlap or intersection over union; which is defined as the size of the intersection of the two pixel sets over the size of their union. 9 To compute the overall quality for the whole database,

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

39

of the two pixel sets over the size of their union. 9 To compute the overall quality for the whole database, we first select the best proposal for each annotated instance with respect to J. The Jaccard index at instance level (Ji) is then defined as the mean best overlap for all the ground-truth instances in the database, also known as Best Spatial Support score (BSS) [4] or Average Best Overlap (ABO) [12]. Computing the mean of the best overlap on all objects, as done by Ji, hides the distribution of quality among different objects. As an example, Ji= 0:5can mean that the algorithm covers half the objects perfectly and completely misses the other half, or can also mean that all the objects are covered exactly at J= 0:5. This information might be useful to decide which algorithm to use. Computing a histogram of the best overlap would provide very rich information, but then the resulting plot would be 3D (number of proposals, bins, and bin counts). Alternatively, we propose to plot different percentiles of the histogram. Interestingly, a certain percentile of the histogram of best overlaps consists in computing the number of objects whose best overlap is above a certain Jaccard

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

40

Interestingly, a certain percentile of the histogram of best overlaps consists in computing the number of objects whose best overlap is above a certain Jaccard threshold, which can be interpreted as the best achievable recall of the technique over a certain threshold. We compute the recall at three different thresholds: J= 0:5, J=0:7, andJ=0:85. Learning Strategy Evaluation : We first estimate the loss in performance due to not sweeping all the possible values offN1;:::;NRgin the combination of proposal lists via the proposed greedy strategy. To do so, we will compare this strategy with the full combination on a reduced problem to make the latter feasible. Specifically, we combine the 4 ranked lists coming from the singletons at all scales, instead of the full 16 lists coming from singletons, pairs, triplets, and 4-tuples. We also limit the search to 20 000 proposals, further speeding the process up. In this situation, the mean loss in achievable quality along the full curve of parameterization is Ji= 0:0002 , with a maximum loss of Ji=0:004(0:74%). In exchange, our proposed learning strategy on the full 16 ranked lists takes about 20 seconds to compute on the training set of

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

41

our proposed learning strategy on the full 16 ranked lists takes about 20 seconds to compute on the training set of SegVOC12, while the singleton-limited full combination takes 4 days (the full combination would take months). Combinatorial Grouping : We now evaluate the Pareto front optimization strategy in the training set of SegVOC12. As before, we extract the lists of proposals from the three scales and the multiscale hierarchy, for singletons, pairs, triplets, and 4-tuples of regions, leading to 16 lists, ranked by the minimum UCM strength of the regions forming each proposal. Figure 8 shows the Pareto front evolution of Jiwith respect to the number of proposals for up to 1, 2, 3, and 4 regions per proposal (4, 8, 12, and 16 lists, respectively) at training and testing time on SegVOC12. As baselines, we plot the raw singletons from MCG-UCM-Our, gPbUCM, and Quadtree; as well as the uniform combination of scales.103104105020406080100 Number of proposalsRegion distribution percentageScale 2.0 Scale 1.0 Scale 0.5 Multi-Scale Fig. 10. Region distribution learnt by the Pareto front optimization on SegVOC12.

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

42

Scale 1.0 Scale 0.5 Multi-Scale Fig. 10. Region distribution learnt by the Pareto front optimization on SegVOC12. The improvement of considering the combination of all 1-region proposals ( ) from the 3 scales and the MCG-UCM-Our with respect to the raw MCG-UCM-Our ( ) is significant, which corroborates the gain in diversity obtained from hierarchies at different scales. In turn, the addition of 2- and 3-region proposals ( and ) noticeably improves the achievable quality. This shows that hierarchies do not get full objects in single regions, which makes sense given that they are built using lowlevel features only. The improvement when adding 4tuples ( ) is marginal at the number of proposals we are considering. When analyzing the equal distribution of proposals from the four scales ( ), we see that the less proposals we consider, the more relevant the Pareto optimization becomes. At the selected working point, the gain of the Pareto optimization is 2 points. Figure 10 shows the distribution of proposals from each of the scales combined in the Pareto front. We see that the coarse scale (0.5) is the most picked at low number of proposals, and the rest come into play when increasing their number, since one can afford more

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

43

see that the coarse scale (0.5) is the most picked at low number of proposals, and the rest come into play when increasing their number, since one can afford more detailed proposals. The multi-scale hierarchy is the one with less weight, since it is created from the other three. Pareto selection and ranking : Back to Figure 8, the red asterisk ( ) marks the selected configuration N1;:::;NR in the Pareto front (black triangle in Figure 7), which is selected at a practical level of proposals. The red plus sign ( ) represents the set of proposals after removing those duplicate proposals whose overlap leads to a Jaccard higher than 0.95. The proposals at this point are the ones that are ranked by the learnt regressor ( ). At test time (right-hand plot), we directly combine the learnt N1;:::;NR proposals from each ranked list. Note that the Pareto optimization does not overfit, given the similar result in the training and validation datasets. We then remove duplicates and rank the results. In this case, note the difference between the regressed result in

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

44

the similar result in the training and validation datasets. We then remove duplicates and rank the results. In this case, note the difference between the regressed result in the training and validation sets, which reflects overfitting, but despite this we found it beneficial with respect to the non-regressed result. 10 Training 1021031041051060:40:50:60:70:80:9 Number of proposalsJaccard index at instance level ( Ji)Validation 1021031041051060:40:50:60:70:80:9 Number of proposalsJaccard index at instance level ( Ji)Pareto up to 4-tuples Pareto up to triplets Pareto up to pairs Pareto only singletons Raw Ours-multi singl. Raw gPb-UCM singl. Raw Quadtree singl. Equal distribution Selected configuration Filtered candidates Regressed ranking Fig. 8. Pareto front evaluation. Achievable quality of our proposals for singletons, pairs, triplets, and 4-tuples; and the raw proposals from the hierarchies on PASCAL SegVOC12 training (left) and validation (right) sets. PASCAL SegVOC12

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

45

the raw proposals from the hierarchies on PASCAL SegVOC12 training (left) and validation (right) sets. PASCAL SegVOC12 1021031040:30:40:50:60:70:80:9 Number of proposalsJaccard index at instance level ( Ji)SBD 1021031040:30:40:50:60:70:80:9 Number of proposalsJaccard index at instance level ( Ji)COCO 1021031040:30:40:50:60:70:80:9 Number of proposalsJaccard index at instance level ( Ji)MCG-Our SCG-Our CPMC [18] CI [19] GOP [13] GLS [14] RIGOR [15] ShSh [17] SeSe [12] MCG-UCM-Our gPb-UCM Quadtree Fig. 9. Object Proposals: Jaccard index at instance level. Results on SegVOC12, SBD, and COCO. In the validation set of SegVOC12, the full set of proposals (i.e., combining the full 16 lists) would contain millions of proposals per image. The multiscale combinatorial grouping allows us to reduce the number of

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

46

proposals (i.e., combining the full 16 lists) would contain millions of proposals per image. The multiscale combinatorial grouping allows us to reduce the number of proposals to 5 086 with a very high achievable Jiof0:81 (). The regressed ranking ( ) allows us to further reduce the number of proposals below this point. Segmented Proposals: Comparison with State of the Art : We first compare our results against those methods that produce segmented object proposals [13], [14], [15], [12], [16], [17], [18], [19], using the implementations from the respective authors. We train MCG on the training set of SegVOC12, and we use the learnt parameters on the validation sets of SegVOC12, SBD, and COCO. Figure 9 shows the achievable quality at instance level (Ji) of all methods on the validation set of SegVOC12, SBD, and COCO. We plot the raw regions of MCGUCM-Our, gPb-UCM, and QuadTree as baselines where available. We also evaluate a faster single-scale version of MCG ( Single-scale Combinatorial Grouping - SCG ), which takes the hierarchy at the native scale only and combines

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

47

MCG ( Single-scale Combinatorial Grouping - SCG ), which takes the hierarchy at the native scale only and combines up to 4 regions per proposal. This approach decreases the computational load one order of magnitude whilekeeping competitive results. MCG proposals ( ) significantly outperform the state-of-the-art at all regimes. The bigger the database is, the better MCG results are with respect to the rest, which shows that our techniques better generalize to unseen images (recall that MCG is trained only in SegVOC12). As commented on the measures description, Jishows mean aggregate results, so they can mask the distribution of quality among objects in the database. Figure 11 shows the recall at three different Jaccard levels. First, these plots further highlight how challenging COCO is, since we observe a significant drop in performance, more pronounced than when measured by JiandJc. Another interesting result comes from observing the evolution of the plots for the three different Jaccard values. Take for instance the performance of GOP ( ) against MCG-Our ( ) in SBD. While for J=0:5GOP slightly outperforms MCG, the higher the threshold, the better MCG. Overall, MCG has specially good results

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

48

slightly outperforms MCG, the higher the threshold, the better MCG. Overall, MCG has specially good results at higherJvalues. In other words, if one looks for proposals of very high accuracy, MCG is the method with highest recall, at all regimes and in all databases. In all measures and databases, SCG ( ) obtains very competitive results, especially if we take into account 11 PASCAL SegVOC12 10210310400:10:20:30:40:50:60:70:80:9 J=0:5 J=0:7 J=0:85 Number of proposalsRecallSBD 10210310400:10:20:30:40:50:60:70:80:9 J= 0:5 J= 0:7 J= 0:85 Number of proposalsRecallCOCO 10210310400:10:20:30:40:50:60:70:80:9 J= 0:5 J= 0:7 J= 0:85 Number of proposalsRecallMCG-Our SCG-Our CPMC [18] CI [19] GOP [13] GLS [14] RIGOR [15] ShSh [17] SeSe [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

49

CI [19] GOP [13] GLS [14] RIGOR [15] ShSh [17] SeSe [12] Quadtree Fig. 11. Segmented Object Proposals: Recall at different Jaccard levels. Percentage of annotated objects for which there is a proposal whose overlap with the segmented ground-truth shapes (not boxes) is above J= 0:5, J= 0:7, andJ= 0:85, for different number of proposals per image. Results on SegVOC12, SBD, and COCO. PASCAL SegVOC12 10210310400:10:20:30:40:50:60:70:80:9 J= 0:5 J= 0:7 J= 0:85 Number of proposalsRecallSBD 10210310400:10:20:30:40:50:60:70:80:9 J= 0:5 J= 0:7 J= 0:85 Number of proposalsRecallCOCO 10210310400:10:20:30:40:50:60:70:80:9 J= 0:5 J= 0:7 J= 0:85 Number of proposalsRecallMCG-Our SCG-Our CPMC [18]

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

50

J= 0:5 J= 0:7 J= 0:85 Number of proposalsRecallMCG-Our SCG-Our CPMC [18] CI [19] GOP [13] GLS [14] RIGOR [15] ShSh [17] SeSe [12] RP [25] EB [26] BING [27] Obj [16] Quadtree Fig. 12. Bounding-Box Proposals: Recall at different Jaccard levels. Percentage of annotated objects for which there is a bounding box proposal whose overlap with the ground-truth boxes is above J= 0:5,J= 0:7, andJ= 0:85, for different number of proposals per image. Results on SegVOC12, SBD, and COCO. that it is 7faster than MCG, as we will see in next sections. The complementarity of MCG with respect to other proposal techniques, and their combination using the Pareto front optimization is studied in [44]. Boxes Proposals: Comparison with State of the Art: Although focused in providing segmented object proposals, MCG may also be used to provide boxes proposals, by taking the bounding box of the segmented

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

51

Art: Although focused in providing segmented object proposals, MCG may also be used to provide boxes proposals, by taking the bounding box of the segmented proposals and deduplicating them. We add the state of the art in boxes proposals [26], [27], [25], and [16] to the comparison. Figure 12 shows the recall values of the boxes results when compared to the bounding boxes of the annotated objects, for three different Jaccard thresholds. While many of the techniques specifically tailored to boxes proposals are competitive at J= 0:5, their perfor-mance drops significantly at higher Jaccard thresholds. Despite being tailored to segmented proposals, MCG clearly outperforms the state of the art if you look for precise localization of the bounding boxes. Again, SCG is very competitive, especially taking its speed into account. MCG and SCG Time per Image : Table 3 shows the time per image of the main steps of our approach, from the raw image to the contours, the hierarchical segmentation, and the proposal generation. All times are computed using a single core on a Linux machine. Our full MCG takes about 25 s. per image to compute the

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

52

segmentation, and the proposal generation. All times are computed using a single core on a Linux machine. Our full MCG takes about 25 s. per image to compute the multiscale hierarchies, and 17 s. to generate and rank the5 038 proposals on the validation set of SegVOC12. Our single-scale SCG produces a segmentation hierarchy of better quality than gPb-ucm [20] in less than 3 seconds and with significant less memory footprint. 12 Fig. 13. COCO Qualitative Results : Image, ground truth, multi-scale UCM and best MCG proposals among the 500 best ranked. (More qualitative examples in the supplemental material.) Contour Hierarchical CandidateTotalDetection Segmentation Generation MCG 4.61.3 20.55.6 17.09.8 42.214.8 SCG 1.00.3 2.70.7 2.60.4 6.21.1 TABLE 3 Time in seconds per image of MCG and SCG Table 4 shows the time-per-image results compared to the rest of state of the art in segmented proposals generation, all run on the same single-core Linux machine. Proposal Generation

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

53

the rest of state of the art in segmented proposals generation, all run on the same single-core Linux machine. Proposal Generation MCG (Our) 42.214.8 SCG (Our) 6.21.1 GOP [13] 1.00.3 GLS [14] 7.91.7 SeSe [12] 15.95.2 RIGOR [15] 31.616.0 CPMC [18] 120 CI [19] 120 ShSh [17] 120 TABLE 4 Time comparison for all considered state-of-the-art techniques that produce segmented proposals. All run on the same single-core Linux machine.Practical considerations : One of the key aspects of object proposals is the size of the pool they generate. Depending on the application, one may need more precision and thus a bigger pool, or one might need speed and thus a small pool in exchange for some loss of quality. MCG and SCG provide a ranked set of around 5 000 and 2 000 proposals, respectively, and one can take the N first in case the specific application needs a smaller pool. From a practical point of view, this means that one does

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

54

first in case the specific application needs a smaller pool. From a practical point of view, this means that one does not need to re-parameterize them for the specific needs of a certain application. In contrast, the techniques that do not provide a ranking of the proposals, need to be re-parameterized to adapt them to a different number of proposals, which is not desirable in practice. On top of that, the results show that MCG and SCG have outstanding generalization power to unseen images (recall that the results for SBD and COCO have been obtained using the learnt parameters on SegVOC12), meaning that MCG and SCG offer the best chance to obtain competitive results in an unseen database without need to re-train. Figure 13 shows some qualitative results on COCO. 7 C ONCLUSIONS We proposed Multiscale Combinatorial Grouping (MCG), a unified approach for bottom-up segmentation and object proposal generation. Our approach produces 13 state-of-the-art contours, hierarchical regions, and object proposals. At its core are a fast eigenvector computation for normalized-cut segmentation and an efficient algorithm for combinatorial merging of hierarchical

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

55

proposals. At its core are a fast eigenvector computation for normalized-cut segmentation and an efficient algorithm for combinatorial merging of hierarchical regions. We also present Single-scale Combinatorial Grouping (SCG), a speeded up version of our technique that produces competitive results in under five seconds per image. We perform an extensive validation in BSDS500, SegVOC12, SBD, and COCO, showing the quality, robustness and scalability of MCG. Recently, an independent study [45], [46] provided further evidence to the interest of MCG among the current state-of-the-art in object proposal generation. Moreover, our object candidates have already been employed as integral part of high performing recognition systems [47]. In order to promote reproducible research on perceptual grouping, all the resources of this project – code, pre-computed results, and evaluation protocols – are publicly available2. Acknowledgements : The last iterations of this work have been done while Jordi Pont-Tuset has been at Prof. Luc Van Gool’s Computer Vision Lab (CVL) of ETHZ, Switzerland. This work has been partly developed in the framework of the project BIGGRAPH-TEC2013-43935-R

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

56

Switzerland. This work has been partly developed in the framework of the project BIGGRAPH-TEC2013-43935-R and the FPU grant AP2008-01164; financed by the Spanish Ministerio de Econom´ ıa y Competitividad , and the European Regional Development Fund (ERDF). This work was partially supported by ONR MURI N000141010933. REFERENCES [1] P . Viola and M. Jones, “Robust real-time face detection,” IJCV , vol. 57, no. 2, 2004. [2] N. Dalal and B. Triggs, “Histograms of oriented gradients for human detection,” in CVPR , 2005. [3] P . Felzenszwalb, R. Girshick, D. McAllester, and D. Ramanan, “Object detection with discriminatively trained part based models,” TP AMI , vol. 32, no. 9, 2010. [4] T. Malisiewicz and A. A. Efros, “Improving spatial support for objects via multiple segmentations,” in BMVC , 2007. [5] C. Gu, J. Lim, P . Arbelaez, and J. Malik, “Recognition using

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

57

[5] C. Gu, J. Lim, P . Arbelaez, and J. Malik, “Recognition using regions,” in CVPR , 2009. [6] J. Carreira, F. Li, and C. Sminchisescu, “Object recognition by sequential figure-ground ranking,” IJCV , vol. 98, no. 3, pp. 243– 262, 2012. [7] A. Ion, J. Carreira, and C. Sminchisescu, “Probabilistic joint image segmentation and labeling by figure-ground composition,” IJCV , vol. 107, no. 1, pp. 40–57, 2014. [8] P . Arbel ´aez, B. Hariharan, C. Gu, S. Gupta, L. Bourdev, and J. Malik, “Semantic segmentation using regions and parts,” in CVPR , 2012. [9] J. Carreira, R. Caseiro, J. Batista, and C. Sminchisescu, “Semantic segmentation with second-order pooling,” in ECCV , 2012.

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

58

segmentation with second-order pooling,” in ECCV , 2012. [10] R. Girshick, J. Donahue, T. Darrell, and J. Malik, “Rich feature hierarchies for accurate object detection and semantic segmentation,” in CVPR , 2014. [11] N. Zhang, J. Donahue, R. Girshick, and T. Darrell, “Part-based r-cnns for fine-grained category detection,” in ECCV , 2014. [12] J. R. R. Uijlings, K. E. A. van de Sande, T. Gevers, and A. W. M. Smeulders, “Selective search for object recognition,” IJCV , vol. 104, no. 2, pp. 154–171, 2013. 2.www.eecs.berkeley.edu/Research/Projects/CS/vision/grouping/mcg/[13] P . Kr ¨ahenb ¨uhl and V . Koltun, “Geodesic object proposals,” in ECCV , 2014. [14] P . Rantalankila, J. Kannala, and E. Rahtu, “Generating 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

59

ECCV , 2014. [14] P . Rantalankila, J. Kannala, and E. Rahtu, “Generating object segmentation proposals using global and local search,” in CVPR , 2014. [15] A. Humayun, F. Li, and J. M. Rehg, “RIGOR: Recycling Inference in Graph Cuts for generating Object Regions,” in CVPR , 2014. [16] B. Alexe, T. Deselaers, and V . Ferrari, “Measuring the objectness of image windows,” TP AMI , vol. 34, pp. 2189–2202, 2012. [17] J. Kim and K. Grauman, “Shape sharing for object segmentation,” inECCV , 2012. [18] J. Carreira and C. Sminchisescu, “CPMC: Automatic object segmentation using constrained parametric min-cuts,” TP AMI , vol. 34, no. 7, pp. 1312–1328, 2012. [19] I. Endres and D. Hoiem, “Category-independent object proposals with diverse ranking,” TP AMI , vol. 36, no. 2, pp. 222–234, 2014.

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

60

with diverse ranking,” TP AMI , vol. 36, no. 2, pp. 222–234, 2014. [20] P . Arbel ´aez, M. Maire, C. C. Fowlkes, and J. Malik, “Contour detection and hierarchical image segmentation,” TP AMI , vol. 33, no. 5, pp. 898–916, 2011. [21] P . F. Felzenszwalb and D. P . Huttenlocher, “Efficient graph-based image segmentation,” IJCV , vol. 59, p. 2004, 2004. [22] X. Ren and L. Bo, “Discriminatively trained sparse code gradients for contour detection,” in NIPS , 2012. [23] C. J. Taylor, “Towards fast and accurate segmentation,” CVPR , 2013. [24] M. Maire and S. X. Yu, “Progressive multigrid eigensolvers for multiscale spectral segmentation,” ICCV , 2013.

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

61

multiscale spectral segmentation,” ICCV , 2013. [25] S. Man ´en, M. Guillaumin, and L. Van Gool, “Prime Object Proposals with Randomized Prim’s Algorithm,” in ICCV , 2013. [26] C. L. Zitnick and P . Doll ´ar, “Edge boxes: Locating object proposals from edges,” in ECCV , 2014. [27] M.-M. Cheng, Z. Zhang, W.-Y. Lin, and P . H. S. Torr, “BING: Binarized normed gradients for objectness estimation at 300fps,” inCVPR , 2014. [28] D. Hoiem, A. Efros, and M. Hebert, “Recovering occlusion boundaries from an image,” IJCV , vol. 91, no. 3, pp. 328–346, 2011. [29] Z. Ren and G. Shakhnarovich, “Image segmentation by cascaded region agglomeration,” in CVPR , 2013. [30] D. Weiss and B. Taskar, “Scalpel: Segmentation cascades with

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

62

region agglomeration,” in CVPR , 2013. [30] D. Weiss and B. Taskar, “Scalpel: Segmentation cascades with localized priors and efficient learning,” in CVPR , 2013. [31] L. Najman and M. Schmitt, “Geodesic saliency of watershed contours and hierarchical segmentation,” TP AMI , vol. 18, no. 12, pp. 1163–1173, 1996. [32] P . Arbel ´aez, “Boundary extraction in natural images using ultrametric contour maps,” in POCV , June 2006. [33] D. Martin, C. Fowlkes, and J. Malik, “Learning to detect natural image boundaries using local brightness, color and texture cues,” TP AMI , vol. 26, no. 5, pp. 530–549, 2004. [34] P . Doll ´ar and C. Zitnick, “Structured forests for fast edge detection,” ICCV , 2013. [35] http://www.eecs.berkeley.edu/Research/ Projects/CS/vision/grouping/resources.html.

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

63

[35] http://www.eecs.berkeley.edu/Research/ Projects/CS/vision/grouping/resources.html. [36] J. Pont-Tuset and F. Marques, “Supervised evaluation of image segmentation and object proposal techniques,” P AMI , 2015. [37] ——, “Measures and meta-measures for the supervised evaluation of image segmentation,” in CVPR , 2013. [38] M. Maire, S. X. Yu, and P . Perona, “Hierarchical scene annotation,” inBMVC , 2013. [39] M. Everingham, H. Muller, and B. Thomas, “Evaluating image segmentation algorithms using the pareto front,” in ECCV , 2006. [40] M. Ehrgott, Multicriteria optimization . Springer, 2005. [41] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman, “The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results,” http://www.pascalnetwork.org/challenges/VOC/voc2012/workshop/index.html.

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

64

Challenge 2012 (VOC2012) Results,” http://www.pascalnetwork.org/challenges/VOC/voc2012/workshop/index.html. [42] B. Hariharan, P . Arbelaez, L. Bourdev, S. Maji, and J. Malik, “Semantic contours from inverse detectors,” in ICCV , 2011. [43] T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P . Perona, D. Ramanan, P . Doll ´ar, and C. Zitnick, “Microsoft COCO: Common Objects in Context,” in ECCV , 2014. [44] J. Pont-Tuset and L. V . Gool, “Boosting object proposals: From pascal to COCO,” in ICCV , 2015. [45] J. Hosang, R. Benenson, and B. Schiele, “How good are detection proposals, really?” in BMVC , 2014. [46] J. Hosang, R. Benenson, P . Doll ´ar, and B. Schiele, “What makes for effective detection proposals?” P AMI , 2015. 14

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

65

for effective detection proposals?” P AMI , 2015. 14 [47] B. Hariharan, P . Arbel ´aez, R. Girshick, and J. Malik, “Simultaneous detection and segmentation,” in ECCV , 2014. Jordi Pont-Tuset is a post-doctoral researcher at ETHZ, Switzerland, in Prof. Luc Van Gool’s computer vision group (2015). He received the degree in Mathematics in 2008, the degree in Electrical Engineering in 2008, the M.Sc. in Research on Information and Communication Technologies in 2010, and the Ph.D with honors in 2014; all from the Universitat Polit `ecnica de Catalunya, BarcelonaTech (UPC). He worked at Disney Research, Z ¨urich (2014). Pablo Arbel ´aezreceived a PhD with honors in Applied Mathematics from the Universit ´e ParisDauphine in 2005. He was a Research Scientist with the Computer Vision Group at UC Berkeley from 2007 to 2014. He currently holds a faculty position at Universidad de los Andes in Colombia. His research interests are in computer vision, where he has worked on a number of problems, including perceptual grouping, object recognition and the analysis of biomedical images.

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

66

Colombia. His research interests are in computer vision, where he has worked on a number of problems, including perceptual grouping, object recognition and the analysis of biomedical images. Jonathan T. Barron is a senior research scientist at Google, working on computer vision and computational photography. He received a PhD in Computer Science from the University of California, Berkeley in 2013, where he was advised by Jitendra Malik, and he received a Honours BSc in Computer Science from the University of Toronto in 2007. His research interests include computer vision, machine learning, computational photography, shape reconstruction, and biological image analysis. He received a National Science Foundation Graduate Research Fellowship in 2009, and the C.V. Ramamoorthy Distinguished Research Award in 2013. Ferran Marques received the degree in Electrical Engineering and the Ph.D. from the Universitat Polit `ecnica de Catalunya, BarcelonaT ech (UPC), where he is currently Professor at the department of Signal Theory and Communications. In the term 2002-2004, he served as President of the European Association for Signal Processing (EURASIP). He has served as Associate Editor of the IEEE Transactions on Image Processing (2009-2012) and as Area Editor for Signal Processing: Image Communication, Elsevier (2010-2014). In 2011, he received the Jaume Vicens

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

67

on Image Processing (2009-2012) and as Area Editor for Signal Processing: Image Communication, Elsevier (2010-2014). In 2011, he received the Jaume Vicens Vives distinction for University Teaching Quality. Currently, he serves as Dean of the Electrical Engineering School (ETSETB-TelecomBCN) at UPC. He has published over 150 conference and journal papers, 2 books, and holds 4 international patents. Jitendra Malik is Arthur J. Chick Professor in the Department of Electrical Engineering and Computer Science at the University of California at Berkeley, where he also holds appointments in vision science and cognitive science. He received the PhD degree in Computer Science from Stanford University in 1985. In January 1986, he joined UC Berkeley as a faculty member in the EECS department where he served as Chair of the Computer Science Division during 2002-2006, and of the Department of EECS during 2004-2006. Jitendra Malik’s group has worked on computer vision, computational modeling of biological vision, computer graphics and machine learning. Several well-known concepts and algorithms arose in this work, such as anisotropic diffusion, normalized cuts, high dynamic range imaging, shape contexts and poselets. According to Google Scholar, ten of his papers have received more than a thousand citations each. He has graduated 33 PhD students. Jitendra was awarded the

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

68

Scholar, ten of his papers have received more than a thousand citations each. He has graduated 33 PhD students. Jitendra was awarded the Longuet-Higgins Award for “A Contribution that has Stood the Test of Time” twice, in 2007 and 2008. He is a Fellow of the IEEE and the ACM, a member of the National Academy of Engineering, and a fellow of the American Academy of Arts and Sciences. He received the PAMI Distinguished Researcher Award in computer vision in 2013 and the K.S. Fu prize in 2014.

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

1504.06201

0

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision Gedas Bertasius University of Pennsylvania gberta@seas.upenn.eduJianbo Shi University of Pennsylvania jshi@seas.upenn.eduLorenzo Torresani Dartmouth College lt@dartmouth.edu Abstract Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting objectlevel features from a pretrained object-classification network. Our method can be viewed as a High-for-Low approach where high-level object features inform the low-level boundary detection process. Our model achieves state-ofthe-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of

1504.06201

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

http://arxiv.org/pdf/1504.06201

cs.CV

20150423

20150921

1504.06201

1

benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of highlevel vision tasks. We demonstrate that by using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a Low-for-High scheme, where lowlevel boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing stateof-the-art high-level vision methods on three distinct tasks. 1. Introduction In the vision community, boundary detection has always been considered a low-level problem. However, psychological studies suggest that when a human observer perceives boundaries, object level reasoning is used [13, 25, 17]. Despite these findings, most of the boundary detection methods rely exclusively on low-level color and gradient features. In this work, we present a method that uses objectlevel features to detect boundaries. We argue that using object-level information to predict boundaries is more sim-Low-Level Task High-Level Tasks BD SBL SS OP ODS MF AP PI-IOU MR

1504.06201

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

http://arxiv.org/pdf/1504.06201

cs.CV

20150423

20150921

1504.06201

2

object-level information to predict boundaries is more sim-Low-Level Task High-Level Tasks BD SBL SS OP ODS MF AP PI-IOU MR SotA 0.76 [27] 28.0 [11] 19.9 [11] 45.8 [5] 0.88 [31] HFL 0.77 62.5 54.6 48.8 0.90 Table 1: Summary of results achieved by our proposed method (H FL) and state-of-the-art methods (SotA). We provide results on four vision tasks: Boundary Detection (BD), Semantic Boundary Labeling (SBL), Semantic Segmentation (SS), and Object Proposal (OP). The evaluation metrics include ODS F-score for BD task, max F-score (MF) and average precision (AP) for SBL task, per image intersection over union (PI-IOU) for SS task, and max recall (MR) for OP task. Our method produces better results in each of these tasks according to these metrics. ilar to how humans reason. Our boundary detection scheme can be viewed as a High-for-Low approach where we use high-level object features as cues for a low-level boundary detection process. Throughout the rest of the paper, we refer to our proposed boundaries as High-for-Low boundaries

1504.06201

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

http://arxiv.org/pdf/1504.06201

cs.CV

20150423

20150921

1504.06201

3

high-level object features as cues for a low-level boundary detection process. Throughout the rest of the paper, we refer to our proposed boundaries as High-for-Low boundaries (HFL). We present an efficient deep network that uses objectlevel information to predict the boundaries. Our proposed architecture reuses features from the sixteen convolutional layers of the network of Simonyan et al. [29], which we refer to as VGG net. The VGG net has been trained for object classification, and therefore, reusing its features allows our method to utilize high-level object information to predict H FL boundaries. In the experimental section, we demonstrate that using object-level features produces semantically meaningful boundaries and also achieves above state-of-the-art boundary detection accuracy. Additionally, we demonstrate that we can successfully apply our H FL boundaries to a number of high-level vision tasks. We show that by using H FL boundaries we improve the results of three existing state-of-the-art methods on the tasks of semantic boundary labeling, semantic segmenta1arXiv:1504.06201v3 [cs.CV] 21 Sep 2015 tion and object proposal generation. Therefore, using H FL boundaries to boost the results in high level vision tasks

1504.06201

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

http://arxiv.org/pdf/1504.06201

cs.CV

20150423

20150921

1504.06201

4

tion and object proposal generation. Therefore, using H FL boundaries to boost the results in high level vision tasks can be viewed as a Low-for-High scheme, where boundaries serve as low-level cues to aid high-level vision tasks. We present the summarized results for the boundary detection and the three mentioned high-level vision tasks in Table 1. Specifically, we compare our proposed method and an appropriate state-of-the-art method for that task. As the results indicate, we achieve better results in each of the tasks for each presented evaluation metric. We present more detailed results for each of these tasks in the later sections. In summary, our contributions are as follows. First, we show that using object-level features for boundary detection produces perceptually informative boundaries that outperform prior state-of-the-art boundary detection methods. Second, we demonstrate that we can use H FL boundaries to enhance the performance on the high-level vision tasks of semantic boundary labeling, semantic segmentation and object proposal. Finally, our method can detect boundaries in near-real time. Thus, we present a boundary detection system that is accurate, efficient, and is also applicable to high level vision tasks. 2. Related Work Most of the contour detection methods predict boundaries based purely on color, text, or other low-level features.

1504.06201

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

http://arxiv.org/pdf/1504.06201

cs.CV

20150423

20150921

1504.06201

5

high level vision tasks. 2. Related Work Most of the contour detection methods predict boundaries based purely on color, text, or other low-level features. We can divide these methods into three broad categories: spectral methods, supervised discriminative methods and deep learning based methods. Spectral methods formulate contour detection problem as an eigenvalue problem. The solution to this problem is then used to reason about the boundaries. The most successful approaches in this genre are the MCG detector [2], gPb detector [1], PMI detector [14], and Normalized Cuts [28]. Some of the notable discriminative boundary detection methods include sketch tokens (ST) [18], structured edges (SE) [6] and sparse code gradients (SCG) [23]. While SCG use supervised SVM learning [4], the latter two methods rely on a random forest classifier and models boundary detection as a classification task. Recently there have been attempts to apply deep learning to the task of boundary detection. SCT [20] is a sparse coding approach that reconstructs an image using a learned dictionary and then detect boundaries. Both N4fields [10] and

1504.06201

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

http://arxiv.org/pdf/1504.06201

cs.CV

20150423

20150921

1504.06201

6

DeepNet [16] approaches use Convolutional Neural Networks (CNNs) to predict edges. N4fields rely on dictionary learning and the use of the Nearest Neighbor algorithm within a CNN framework while DeepNet uses a traditional CNN architecture to predict contours. The most similar to our approach is DeepEdge [3], which uses a multi-scale bifurcated network to perform contour detection using object-level features. However, we show that our method achieves better results even without 1100x1100 512 500x375... 1024 Input Image 5504 Convolutional Feature Maps 500x375Upsampled Image 5504-Dimensional Feature Vectors Predicted Boundaries |{z} Feature Interpolation & Concatenation 1100x1100x64- Candidate Contour Points 35x35x512 Shared Weights for all Candidate Points |{z} Figure 1: An illustration of our architecture (best viewed in color). First we extract a set of candidate contour points. Then we upsample the image and feed it through 16convolutional layers pretrained for object classification. For each candidate point, we find its correspondence in each of the feature maps and perform feature interpolation. This yields

1504.06201

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

http://arxiv.org/pdf/1504.06201

cs.CV

20150423

20150921

1504.06201

7

candidate point, we find its correspondence in each of the feature maps and perform feature interpolation. This yields a5504 -dimensional feature vector for each candidate point. We feed each of these vectors to two fully connected layers and store the predictions to produce a final boundary map. the complicated multi-scale and bifurcated architecture of DeepEdge. Additionally, unlike DeepEdge, our system can run in near-real time. In comparison to prior approaches, we offer several contributions. First, we propose to use object-level information to predict boundaries. We argue that such an approach leads to semantic boundaries, which are more consistent with humans reasoning. Second, we avoid feature engineering by learning boundaries from human-annotated data. Finally, we demonstrate excellent results for both low-level and high-level vision tasks. For the boundary detection task, our proposed H FL boundaries outperform all of the prior methods according to both F-score metrics. Additionally, we show that because H FL boundaries are based on objectlevel features, they can be used to improve performance in the high-level vision tasks of semantic boundary labeling, semantic segmentation, and object proposal generation. 3. Boundary Detection In this section, we describe our proposed architecture and the specific details on how we predict H FL boundaries

1504.06201

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

http://arxiv.org/pdf/1504.06201

cs.CV

20150423

20150921

1504.06201

8

3. Boundary Detection In this section, we describe our proposed architecture and the specific details on how we predict H FL boundaries using our method. The detailed illustration of our architecture is presented in Figure 1. Figure 2: A visualization of selected convolutional feature maps from VGG network (resized to the input image dimension). Because VGG was optimized for an object classification task, it produces high activation values on objects and their parts. 3.1. Selection of Candidate Contour Points We first extract a set of candidate contour points with a high recall. Due to its efficiency and high recall performance, we use the SE edge detector [6]. In practice, we could eliminate this step and simply try to predict boundaries at every pixel. However, selecting a set of initial candidate contour points, greatly reduces the computational cost. Since our goal is to build a boundary detector that is both accurate and efficient, we use these candidate points to speed up the computation of our method. 3.2. Object-Level Features After selecting candidate contour points, we up-sample the original input image to a larger dimension (for example 11001100 ). The up-sampling is done to minimize the loss of information due to the input shrinkage caused by pooling

1504.06201

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

http://arxiv.org/pdf/1504.06201

cs.CV

20150423

20150921

1504.06201

9

the original input image to a larger dimension (for example 11001100 ). The up-sampling is done to minimize the loss of information due to the input shrinkage caused by pooling at the different layers. Afterwards, we feed the up-sampled image through 16convolutional layers of the VGG net. We use the VGG net as our model because it has been trained to recognize a large number of object classes (the 1000 categories of the ImageNet dataset [24]) and thus encodes object-level features that apply to many classes. To preserve specific location information we utilize only the 16convolutional layers of the VGG net. We don’t use fully connected layers because they don’t preserve spatial information, which is crucial for accurate boundary detection. We visualize some of the selected convolutional maps in Figure 2. Note the high activation values around the various objects in the images, which confirms our hypothesis that the VGG net encodes object specific information in its convolutional feature maps. 3.3. Feature Interpolation Similarly to [26, 12, 19], we perform feature interpolation in deep layers. After the up-sampled image passes

1504.06201

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

http://arxiv.org/pdf/1504.06201

cs.CV

20150423

20150921

1504.06201

10

3.3. Feature Interpolation Similarly to [26, 12, 19], we perform feature interpolation in deep layers. After the up-sampled image passes through all 16convolutional layers, for each selected candidate contour point we find its corresponding point in the feature maps. Due to the dimension differences in convo-lutional maps these correspondences are not exact. Thus we perform feature interpolation by finding the four nearest points and averaging their activation values. This is done in each of the 5504 feature maps. Thus, this results in a 5504 -dimensional vector for each candidate point. We note that the interpolation of convolutional feature maps is the crucial component that enables our system to predict the boundaries efficiently. Without feature interpolation, our method would need to independently process the candidate edge points by analyzing a small image patch around each point, as for example done in DeepEdge [3] which feeds one patch at a time through a deep network. However, when the number of candidate points is large (e.g., DeepEdge considers about 15K points at each of 4 different scales), their patches overlap significantly and thus a large amount of computation is wasted by recalculating filter response values over the same pixels. Instead, we can

1504.06201

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

http://arxiv.org/pdf/1504.06201

cs.CV

20150423

20150921

1504.06201

11

large amount of computation is wasted by recalculating filter response values over the same pixels. Instead, we can compute the features for all candidate points with a single pass through the network by performing deep convolution over the entire image (i.e., feeding the entire image rather than one patch at a time) and then by interpolating the convolutional feature maps at the location of each candidate edge point so as to produce its feature descriptor. Thanks to this speedup, our method has a runtime of 1:2seconds (using a K40 GPU), which is better than the runtimes of prior deep-learning based edge detection methods [27, 10, 16, 3]. 3.4. Learning to Predict Boundaries After performing feature interpolation, we feed the 5504 -dimensional feature vectors corresponding to each of the candidate contour points to two fully connected layers that are optimized to the human agreement criterion. To be more precise, we define our prediction objective as a fraction of human annotators agreeing on the presence of the boundary at a particular pixel. Therefore, a learning objective aims at mimicking the judgement of the human labelers. Finally, to detect H FL boundaries, we accumulate the predictions from the fully connected layers for each of the candidate points and produce a boundary probability map as illustrated in Figure 1.

1504.06201

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

http://arxiv.org/pdf/1504.06201

cs.CV

20150423

20150921

1504.06201

12

Finally, to detect H FL boundaries, we accumulate the predictions from the fully connected layers for each of the candidate points and produce a boundary probability map as illustrated in Figure 1. 3.5. Implementation Details In this section, we describe the details behind the training procedure of our model. We use the Caffe library [15] to implement our network architecture. In the training stage, we freeze the weights in all of the convolutional layers. To learn the weights in the two fully connected layers we train our model to optimize the least squares error of the regression criterion that we described in the previous subsection. To enforce regularization we set a dropout rate of 0:5in the fully connected layers. Our training dataset includes 80Kpoints from the BSDS500 dataset [22]. As described in the previous subsection, the labels represent the fraction of human annotators agreeing on the boundary presence. We divide the label space into four quartiles, and select an equal number of samples for each quartile to balance the training dataset. In addition to the training dataset, we also sample a hold-out dataset of size 40;000. We use this for the hard-positive mining [21] in order to reduce the number of false-negative predictions.

1504.06201

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

http://arxiv.org/pdf/1504.06201

cs.CV

20150423

20150921

1504.06201

13

dataset of size 40;000. We use this for the hard-positive mining [21] in order to reduce the number of false-negative predictions. For the first 25epochs we train the network on the original80;000training samples. After the first 25epochs, we test the network on the hold-out dataset and detect false negative predictions made by our network. We then augment the original 80;000 training samples with the false negatives and the same number of randomly selected true negatives. For the remaining 25epochs, we train the network on this augmented dataset. 3.6. Boundary Detection Results In this section, we present our results on the BSDS500 dataset [22], which is the most established benchmark for boundary detection. The quality of the predicted boundaries is evaluated using three standard measures: fixed contour threshold (ODS), per-image best threshold (OIS), and average precision (AP). We compare our approach to the state-of-the-art based on two different sets of BSDS500 ground truth boundaries. First, we evaluate the accuracy by matching each of the predicted boundary pixels with the ground truth boundaries that were annotated by anyof the human annotators. This

1504.06201

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

http://arxiv.org/pdf/1504.06201

cs.CV

20150423

20150921

1504.06201

14

First, we evaluate the accuracy by matching each of the predicted boundary pixels with the ground truth boundaries that were annotated by anyof the human annotators. This set of ground truth boundaries is referred to as “any”. We present the results for “any” ground truth boundaries in the lower half of Table 2. As indicated by the results, H FL boundaries outperform all the prior methods according to both F-score measures. Recently, there has been some criticism raised about the procedure for boundary detection evaluation on the BSDS500 dataset. One issue with the BSDS500 dataset involves the so called “orphan” boundaries: the bound-Consensus GT ODS OIS AP FPS SCG [23] 0.6 0.64 0.56 1/280 DeepNet [16] 0.61 0.64 0.61 1=5z PMI [14] 0.61 0.68 0.56 1/900 DeepEdge [3] 0.62 0.64 0.64 1=1000z N4-fields [10] 0.64 0.67 0.64 1=6z HFL 0.65 0.68 0.67 5=6z Any GT ODS OIS AP FPS

1504.06201

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

http://arxiv.org/pdf/1504.06201

cs.CV

20150423

20150921

1504.06201

15

HFL 0.65 0.68 0.67 5=6z Any GT ODS OIS AP FPS SE [6] 0.75 0.77 0.80 2.5 MCG [2] 0.75 0.78 0.76 1/24 N4-fields [10] 0.75 0.77 0.78 1=6z DeepEdge [3] 0.75 0.77 0.81 1=1000z MSC [30] 0.76 0.78 0.79 DeepContour [27] 0.76 0.77 0.8 1=30z HFL 0.77 0.79 0.8 5=6z Table 2: Boundary detection results on BSDS500 benchmark. Upper half of the table illustrates the results for “consensus” ground-truth criterion while the lower half of the table depicts the results for “any” ground-truth criterion. In both cases, our method outperforms all prior methods according to both ODS (optimal dataset scale) and OIS (optimal image scale) metrics. We also report the run-time of our method (zGPU time) in the FPS column (frames per second), which shows that our algorithm is faster than prior approaches based on deep learning [27, 10, 16, 3].

1504.06201

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

http://arxiv.org/pdf/1504.06201

cs.CV

20150423

20150921

1504.06201

16

second), which shows that our algorithm is faster than prior approaches based on deep learning [27, 10, 16, 3]. Figure 5: Qualitative results on the BSDS benchmark. The first column of images represent input images. The second column illustrates SE [6], while the third column depicts HFL boundaries. Notice that SE boundaries are predicted with low confidence if there is no significant change in color between the object and the background. Instead, because our model is defined in terms of object-level features, it can predict object boundaries with high confidence even if there is no significant color variation in the scene. aries that are marked by only one or two human annotators. These “orphan” boundaries comprise around 30% of BSDS500 dataset but most of them are considered uninformative. However, the standard evaluation benchmark rewards the methods that predict these boundaries. To resolve 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 5500024x 10−3Weight Magnitude Feature Map Number (Sorted in the Order from Early to Deep Layers)Figure 6: We train a linear regression model and visualize its weight magnitudes in order to understand which features are used most heavily in the boundary prediction (this linear

1504.06201

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

http://arxiv.org/pdf/1504.06201

cs.CV

20150423

20150921

1504.06201

17

its weight magnitudes in order to understand which features are used most heavily in the boundary prediction (this linear regression is used only for the visualization purposes and not for the accuracy analysis). Note how heavily weighted features lie in the deepest layers of the network, i.e., the layers that are most closely associated with object information. this issue we also evaluate our H FL boundaries on the so called “consensus” set of ground truth boundaries. These “consensus” boundaries involve only boundaries that are marked by allof the human annotators and hence, are considered perceptually meaningful. In the upper half of Table 2, we present the results achieved by our method on the “consensus” set of the ground truth boundaries. Our H FL boundaries outperform or tie all the prior methods in each of the three evaluation metrics, thus suggesting that H FL boundaries are similar to the boundaries that humans annotated. We also report the runtimes in Table 2 and note that our method runs faster than previous deep-learning based edge detection systems [27, 10, 16, 3]. Our proposed model computes a highly nonlinear function of the 5504-dimensional feature vector of each candidate point. Thus, it is difficult to assess which features are used most heavily by our edge predictor. However, we can

1504.06201

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

http://arxiv.org/pdf/1504.06201

cs.CV

20150423

20150921

1504.06201

18

used most heavily by our edge predictor. However, we can gain a better insight by replacing the nonlinear function with a simple linear model. In Fig. 6 we show the weight magnitudes of a simple linear regression model (we stress that this linear model is used only for feature visualization purposes). From this Figure, we observe that many important features are located in the deepest layers of the VGG network. As shown in [7], these layers encode high-level object information, which confirms our hypothesis that highlevel information is useful for boundary detection. Finally, we present some qualitative results achieved by our method in Figure 5. These examples illustrate the effective advantage that H FL boundaries provide over another state-of-the-art edge detection system, the SE system [6]. Specifically, observe the parts of the image where there is a boundary that separates an object from the background but where the color change is pretty small. Notice that because the SE boundary detection is based on low-level color and texture features, it captures these boundaries with very low confidence. In comparison, because H FL boundaries rely on object-level features, it detects these boundaries with high confidence.4. High-Level Vision Applications In this section, we describe our proposed Low-for-High

1504.06201

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

http://arxiv.org/pdf/1504.06201

cs.CV

20150423

20150921

1504.06201

19

high confidence.4. High-Level Vision Applications In this section, we describe our proposed Low-for-High pipeline: using low-level boundaries to aid a number of high-level vision tasks. We focus on the tasks of semantic boundary labeling, semantic segmentation and object proposal generation. We show that using H FL boundaries improves the performance of state-of-the-art methods in each of these high-level vision tasks. 4.1. Semantic Boundary Labeling The task of semantic boundary labeling requires not only to predict the boundaries but also to associate a specific object class to each of the boundaries. This implies that given our predicted boundaries we also need to label them with object class information. We approach this problem by adopting the ideas from the recent work on Fully Convolutional Networks (FCN) [19]. Given an input image, we concurrently feed it to our boundary-predicting network (described in Section 3), and also through the FCN that was pretrained for 20Pascal VOC classes and the background class. While our proposed network produces H FL boundaries, the FCN model predicts class probabilities for each of the pixels. We can then merge the two output maps as follows. For a given boundary point we consider a 99

1504.06201

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

http://arxiv.org/pdf/1504.06201

cs.CV

20150423

20150921

1504.06201

20

of the pixels. We can then merge the two output maps as follows. For a given boundary point we consider a 99 grid around that point from each of the 21FCN object-class probability maps. We calculate the maximum value inside each grid, and then label the boundary at a given pixel with the object-class that corresponds to the maximum probability across these 21maps. We apply this procedure for each of the boundary points, in order to associate object-class labels to the boundaries. Note that we consider the grids around the boundary pixel because the output of the FCN has a poor localization, and considering the grids rather than individual pixels leads to higher accuracy. We can also merge H FL boundaries with the state-ofthe-art DeepLab-CRF segmentation [5] to obtain higher accuracy. We do this in a similar fashion as just described. First, around a given boundary point we extract a 99 grid from the DeepLab-CRF segmentation. We then compute the mode value in the grid (excluding the background class), and use the object-class corresponding to the mode value as a label for the given boundary point. We do this for each of the boundary points. By merging H FL boundaries and the output of FCN or DeepLab-CRF, we get semantic

1504.06201

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

http://arxiv.org/pdf/1504.06201

cs.CV

20150423

20150921

1504.06201

21

each of the boundary points. By merging H FL boundaries and the output of FCN or DeepLab-CRF, we get semantic boundaries that are highly localized and also contain objectspecific information. 4.1.1 Semantic Boundary Labeling Results In this section, we present semantic boundary labeling results on the SBD dataset [11], which includes ground truth boundaries that are also labeled with one of 20Pascal VOC classes. The boundary detection accuracy for each class is Method (Metric) aero bike bird boat bottle bus car cat chair cow table dog horse mbike person plant sheep sofa train tv mean InvDet (MF) 42.6 49.5 15.7 16.8 36.7 43.0 40.8 22.6 18.1 26.6 10.2 18.0 35.2 29.4 48.2 14.3 26.8 11.2 22.2 32.0 28.0 HFL-FC8 (MF) 71.6 59.6 68.0 54.1 57.2 68.0 58.8 69.3 43.3 65.8 33.3 67.9 67.5 62.2 69.0 43.8 68.5 33.9 57.7 54.8 58.7

1504.06201

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

http://arxiv.org/pdf/1504.06201

cs.CV

20150423

20150921

1504.06201

22

HFL-CRF (MF) 73.9 61.4 74.6 57.2 58.8 70.4 61.6 71.9 46.5 72.3 36.2 71.1 73.0 68.1 70.3 44.4 73.2 42.6 62.4 60.1 62.5 InvDet (AP) 38.4 29.6 9.6 9.9 24.2 33.6 31.3 17.3 10.7 16.4 3.7 12.1 28.5 20.4 45.7 7.6 16.1 5.7 14.6 22.7 19.9 HFL-FC8 (AP) 66.0 50.7 58.9 40.6 47.1 62.9 51.0 59.0 25.6 54.6 15.3 57.8 57.3 55.9 62.2 27.5 55.6 18.0 50.1 40.6 47.8 HFL-CRF (AP) 71.2 55.2 69.3 45.7 48.9 71.1 56.8 65.7 29.1 65.9 17.7 64.5 68.3 64.7 65.9 29.1 66.5 25.7 60.0 49.8 54.6

1504.06201

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

http://arxiv.org/pdf/1504.06201

cs.CV

20150423

20150921

1504.06201

23

Table 3: Results of semantic boundary labeling on the SBD benchmark using the Max F-score (MF) and Average Precision (AP) metrics. Our method (H FL) outperforms Inverse Detectors [11] for all 20categories according to both metrics. Note that using the CRF output to label the boundaries produces better results than using the outputs from the FC8 layer of FCN. evaluated using the maximum F-score (MF), and average precision (AP) measures. Labeling boundaries with the semantic object information is a novel and still relatively unexplored problem. Therefore, we found only one other approach (Inverse Detectors) that tried to tackle this problem [11]. The basic idea behind Inverse Detectors consists of several steps. First, generic boundaries in the image are detected. Then, a number of object proposal boxes are generated. These two sources of information are then used to construct the features. Finally, a separate classifier is used to label the boundaries with the object-specific information. Table 3 shows that our approach significantly outperforms Inverse Detectors according to both the maximum Fscore and the average precision metrics for all twenty categories. As described in Section 4.1 we evaluate the two

1504.06201

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

http://arxiv.org/pdf/1504.06201

cs.CV

20150423

20150921

1504.06201

24

variants of our method. Denoted by H FL-FC8 is the variant for which we label H FL boundaries with the outputs from the last layer (FC8) of the pretrained FCN. We denote with H FL-CRF the result of labeling our boundaries with the output from the DeepLab-CRF [5]. Among these two variants, we show that the latter one produces better results. This is expected since the CRF framework enforces spatial coherence in the semantic segments. In Figure 7, we present some of the qualitative results produced by our method. We note that even with multiple objects in the image, our method successfully recognizes and localizes boundaries of each of the classes. 4.2. Semantic Segmentation For the semantic segmentation task, we propose to enhance the DeepLab-CRF [5] with our predicted H FL boundaries. DeepLab-CRF is a system comprised of a Fully Convolutional Network (described in Section 4.1) and a dense CRF applied on top of FCN predictions. Specifically, in the CRF, the authors propose to use a Gaussian kernel and a bilateral term including position and color terms as the CRF features (see [5]). While in most cases the proposed scheme works well, DeepLab-CRF

1504.06201

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

http://arxiv.org/pdf/1504.06201

cs.CV

20150423

20150921

1504.06201

25

a Gaussian kernel and a bilateral term including position and color terms as the CRF features (see [5]). While in most cases the proposed scheme works well, DeepLab-CRF sometimes produces segmentations that are not spatially coherent, particularly for images containing small object reFigure 7: A visualization of the predicted semantic boundary labels. Images in the first column are input examples. Columns two and three show semantic H FL boundaries of different object classes. Note that even with multiple objects appearing simultaneously, our method outputs precise semantic boundaries. gions. We propose to address this issue by adding features based on our predicted H FL boundaries in the CRF framework. Note that we use predicted boundaries from Section 3 and not the boundaries labeled with the object information that we obtained in Section 4.1. We use the Normalized Cut [28] framework to generate our features. First, we construct a pixel-wise affinity matrix Wusing our H FL boundaries. We measure the similarity between two pixels as: Wij= exp (max p2ijfM(p)2 2g) whereWijrepresents the similarity between pixels iand j,pdenotes the boundary point along the line segment ij connecting pixels iandj,Mdepicts the magnitude of the

1504.06201

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

http://arxiv.org/pdf/1504.06201

cs.CV

20150423

20150921

1504.06201

26

j,pdenotes the boundary point along the line segment ij connecting pixels iandj,Mdepicts the magnitude of the boundary at pixel p, anddenotes the smoothness parameter, which is usually set to 14% of the maximum boundary value in the image. The intuitive idea is that two pixels are similar (i.e. Wij= 1) if there is no boundary crossing the line connecting these two pixels (i.e. M(p) = 08p2ij) or if the boundary strength is low. We note that it is not necessary to build a full affinity matrix W. We build a sparse affinMetric Method (Dataset) aero bike bird boat bottle bus car cat chair cow table dog horse mbike person plant sheep sofa train tv mean PP-IOUDL-CRF (VOC) 78.6 41.1 83.5 75.3 72.9 83.1 76.6 80.8 37.8 72.1 66.5 64.7 65.8 75.7 80.5 34.4 75.9 47.4 86.6 77.9 68.9

1504.06201

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

http://arxiv.org/pdf/1504.06201

cs.CV

20150423

20150921

1504.06201

27

DL-CRF+H FL (VOC) 77.9 41.2 83.1 74.4 73.2 85.5 76.1 80.6 35.7 71.0 66.6 64.3 65.9 75.2 80.2 32.8 75.2 47.0 87.1 77.9 68.5 DL-CRF (SBD) 74.2 68.0 81.9 64.6 71.8 86.3 78.3 84.3 41.6 78.0 49.9 82.0 78.5 77.1 80.1 54.3 75.6 49.8 79.5 70.1 71.4 DL-CRF+H FL (SBD) 75.1 69.2 81.6 64.8 71.3 86.4 78.1 84.1 41.2 77.8 50.4 81.6 78.2 78.5 80.7 53.8 74.9 49.1 79.5 70.4 71.4 PI-IOUDL-CRF (VOC) 46.1 28.0 48.5 54.5 45.5 57.6 34.1 47.3 19.5 61.4 41.6 42.5 34.4 61.8 62.1 22.1 50.5 41.0 61.2 31.9 44.6

1504.06201

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

http://arxiv.org/pdf/1504.06201

cs.CV

20150423

20150921

1504.06201

28

DL-CRF+H FL (VOC) 47.5 27.6 50.4 63.5 47.7 57.9 38.7 47.2 21.1 57.3 41.2 43.7 36.0 66.4 61.1 21.3 53.9 42.1 70.9 34.6 46.5 DL-CRF (SBD) 59.4 36.5 58.0 38.6 32.0 58.1 44.7 59.6 25.8 51.8 28.1 59.0 46.9 50.3 61.8 22.2 45.9 33.4 62.1 41.0 45.8 DL-CRF+H FL (SBD) 63.4 42.5 58.4 41.3 32.5 61.2 45.7 61.4 28.4 55.5 31.5 61.4 51.8 54.6 62.1 24.9 52.6 34.2 67.1 45.1 48.8 Table 4: Semantic segmentation results on the SBD and VOC 2007 datasets. We measure the results according to PP-IOU (per pixel) and PI-IOU (per image) evaluation metrics. We denote the original DeepLab-CRF system and our proposed

1504.06201

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

http://arxiv.org/pdf/1504.06201

cs.CV

20150423

20150921

1504.06201

29

(per pixel) and PI-IOU (per image) evaluation metrics. We denote the original DeepLab-CRF system and our proposed modification as DL-CRF and DL-CRF+H FL, respectively. According to the PP-IOU metric, our proposed features (DLCRF+H FL) yield almost equivalent results as the original DeepLab-CRF system. However, based on PI-IOU metric, our proposed features improve the mean accuracy by 3%and1:9%on SBD and VOC 2007 datasets respectively. ity matrix connecting every pair of pixels iandjthat have distance 5or less from each other. After building a boundary-based affinity matrix Wwe setDii=P i6=jWijand compute eigenvectors vof the generalized eigenvalue system: (DW)v=Dv We then resize the eigenvectors vto the original image dimensions, and use them as additional features to the CRF part of DeepLab-CRF system. In our experiments, we use the16eigenvectors corresponding to the smallest eigenvalues, which results in 16extra feature channels. Note that the eigenvectors contain soft segmentation information. Because H FL boundaries predict object-level

1504.06201

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

http://arxiv.org/pdf/1504.06201

cs.CV

20150423

20150921

1504.06201

30

Note that the eigenvectors contain soft segmentation information. Because H FL boundaries predict object-level contours with high confidence, the eigenvectors often capture regions corresponding to objects. We visualize a few selected eigenvectors in Figure 8. In the experimental section, we demonstrate that our proposed features make the output produced by DeepLab-CRF more spatially coherent and improve the segmentation accuracy according to one of the metrics. We also note that our proposed features are applicable to any generic method that incorporates CRF. For instance, even if DeepLab-CRF used an improved DeepLab network architecture, our features would still be beneficial because they contribute directly to the CRF part and not the DeepLab network part of the system. 4.2.1 Semantic Segmentation Results In this section, we present semantic segmentation results on the SBD [11] and also Pascal VOC 2007 [8] datasets, which both provide ground truth segmentations for 20Pascal VOC classes. We evaluate the results in terms of two metrics. The first metric measures the accuracy in terms of pixel intersection-over-union averaged per pixel (PP-IOU) across the 20 classes. According to this metric, the accuracy is computed on a per pixel basis. As a result, the images that

1504.06201

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

http://arxiv.org/pdf/1504.06201

cs.CV

20150423

20150921

1504.06201

31

across the 20 classes. According to this metric, the accuracy is computed on a per pixel basis. As a result, the images that contain large object regions are given more importance. Figure 8: In this figure, the first column depicts an input image while the second and third columns illustrate two selected eigenvectors for that image. The eigenvectors contain soft segmentation information. Because H FL boundaries capture object-level boundaries, the resulting eigenvectors primarily segment regions corresponding to the objects. We observe that while DeepLab-CRF works well on the images containing large object regions, it produces spatially disjoint outputs for the images with smaller and object regions (see Figure 9). This issue is often being overlooked, because according to the PP-IOU metric, the images with large object regions are given more importance and thus contribute more to the accuracy. However, certain applications may require accurate segmentation of small objects. Therefore, in addition to PP-IOU, we also consider the PIIOU metric (pixel intersection-over-union averaged per image across the 20 classes), which gives equal weight to each of the images. For both of the metrics we compare the semantic segmentation results of a pure DeepLab-CRF [5] and also a

1504.06201

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

http://arxiv.org/pdf/1504.06201

cs.CV

20150423

20150921

1504.06201

32

of the images. For both of the metrics we compare the semantic segmentation results of a pure DeepLab-CRF [5] and also a modification of DeepLab-CRF with our proposed features added to the CRF framework. We present the results for both of the metrics in Table 4. Based on these results, we observe that according to the first metric (PP-IOU), our proposed features yield almost equivalent results as the original DeepLab-CRF system. However, according to the second metric (PI-IOU) our features yield an average improvement of 3%and1:9%in Figure 9: An illustration of the more challenging semantic segmentation examples. The first column depicts the predictions achieved by DeepLab-CRF, while the second column illustrates the results after adding our proposed features to the CRF framework. The last column represents ground truth segmentations. Notice how our proposed features render the predicted semantic segments more spatially coherent and overall more accurate. SBD and VOC 2007 datasets respectively. We also visualize the qualitative results produced by both approaches in Figure 9. Notice how our proposed features make the segmentations look smoother relative to the segmentations produced by the original DeepLab-CRF system. Once again, we want to stress that our H FL features

1504.06201

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

http://arxiv.org/pdf/1504.06201

cs.CV

20150423

20150921

1504.06201

33

make the segmentations look smoother relative to the segmentations produced by the original DeepLab-CRF system. Once again, we want to stress that our H FL features are applicable to any method that uses the CRF. Therefore, based on the results presented in this section, we believe that our proposed features could be beneficial in a wide array of problems that involve the use of the CRF framework. 4.3. Object Proposals Finally, we show that our method produces object-level boundaries that can be successfully exploited in an object proposal scheme. Specifically we adopt the EdgeBoxes approach [31], which can be applied to any generic boundaries to produce a list of object proposal boxes. The original EdgeBoxes method uses SE boundaries to generate the boxes. However, SE boundaries are predicted using lowlevel color and texture features, rather than object-level features. Thus, here we validate the hypothesis that the EdgeBoxes proposals can be improved by replacing the SE boundaries with our H FL boundaries. 4.3.1 Object Proposal Results In this section, we present object proposal results on the Pascal VOC 2012 dataset [9]. We evaluate the quality of bounding-box proposals according to three metrics: area under the curve (AUC), the number of proposals needed to

1504.06201

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

http://arxiv.org/pdf/1504.06201

cs.CV

20150423

20150921

1504.06201

34

Pascal VOC 2012 dataset [9]. We evaluate the quality of bounding-box proposals according to three metrics: area under the curve (AUC), the number of proposals needed to reach recall of 75%, and the maximum recall over 5000 object bounding-boxes. Additionally, we compute the accuracy for each of the metrics for three different intersectionMethodIoU 0.65 IoU 0.7 IoU 0.75 AUC N@75% Recall AUC N@75% Recall AUC N@75% Recall SE 0:52 413 0.93 0:47 658 0.88 0.41 inf 0:75 HFL 0.53 365 0.95 0.48 583 0.9 0.41 2685 0.77 Table 5: Comparison of object proposal results. We compare the quality of object proposals using Structured Edges [6] and H FL boundaries. We evaluate the performance for three different IOU values and demonstrate that using H FL boundaries produces better results for each evaluation metric and for each IOU value. over union (IOU) values: 0:65;0:7, and 0:75. We present these results in Table 5. As described in Section 4.3, we use EdgeBoxes [31], a package that uses generic boundaries, to generate object proposals. We compare the quality of the

1504.06201

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

http://arxiv.org/pdf/1504.06201

cs.CV

20150423

20150921

1504.06201

35

EdgeBoxes [31], a package that uses generic boundaries, to generate object proposals. We compare the quality of the generated object proposals when using SE boundaries and HFL boundaries. We demonstrate that for each IOU value and for each of the three evaluation metrics, H FL boundaries produce better or equivalent results. This confirms our hypothesis that H FL boundaries can be used effectively for high-level vision tasks such as generating object proposals. 5. Conclusions In this work, we presented an efficient architecture that uses object-level information to predict semantically meaningful boundaries. Most prior edge detection methods rely exclusively on low-level features, such as color or texture, to detect the boundaries. However, perception studies suggest that humans employ object-level reasoning when deciding whether a given pixel is a boundary [13, 25, 17]. Thus, we propose a system that focuses on the semantic objectlevel cues rather than low level image information to detect the boundaries. For this reason we refer to our boundary detection scheme as a High-for-Low approach, where highlevel object features inform the low-level boundary detection process. In this paper we demonstrated that our proposed method produces boundaries that accurately separate objects and the background in the image and also achieve higher F-score compared to any prior work.

1504.06201

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

http://arxiv.org/pdf/1504.06201

cs.CV

20150423

20150921

1504.06201

36

objects and the background in the image and also achieve higher F-score compared to any prior work. Additionally, we showed that, because H FL boundaries are based on object-level features, they can be employed to aid a number of high level vision tasks in a Low-for-High fashion. We use our boundaries to boost the accuracy of state-of-the-art methods on the high-level vision tasks of semantic boundary labeling, semantic segmentation, and object proposals generation. We show that using H FL boundaries leads to better results in each of these tasks. To conclude, our boundary detection method is accurate, efficient, applicable to a variety of datasets, and also useful for multiple high-level vision tasks. We plan to release the source code for H FL upon the publication of the paper . 6. Acknowledgements We thank Mohammad Haris Baig for the suggestions and help with the software. This research was funded in part by NSF award CNS-1205521. References [1] P. Arbelaez, M. Maire, C. Fowlkes, and J. Malik. Contour detection and hierarchical image segmentation. IEEE Trans. Pattern Anal. Mach. Intell. , 33(5):898–916, May 2011. 2

1504.06201

High-for-Low and Low-for-High: Efficient Boundary Detection from Deep Object Features and its Applications to High-Level Vision

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

http://arxiv.org/pdf/1504.06201

cs.CV

20150423

20150921