Convolutional Oriented Boundaries

1 Introduction

• COB learns multiscale oriented contours with a generic CNN architecture and uses a sparse boundary representation to construct hierarchical regions efficiently.Figure 1 depicts contours from one CNN pass, scale-specific Ultrametric Contour Maps, and their fusion into a hierarchical segmentation structure.Source: §1 · p. 1, Figure 1 · p. 2
• The BSDS has only 300 training images and 200 test images, where leading boundary methods are nearing human performance, motivating evaluation on larger, harder data.PASCAL VOC trainval has more than 10,000 varied images; PASCAL Context extends boundary annotations to background categories and provides fully parsed images.Source: §1 · p. 2
• COB achieves state-of-the-art contours and regions on PASCAL Context and BSDS, runs in 0.8 seconds per image, and reports strong object-proposal results on PASCAL Segmentation and MS-COCO.The MS-COCO proposals are evaluated without retraining and show a large improvement over the state of the art.Source: §1 · pp. 1–3
• The COB code, precomputed results, pretrained models, and benchmarks are publicly available.Source: §1 · p. 3

2 Related Work

• Deep contour methods use patch classification, CNN-feature regression, shape-informed features, or convolutional feature maps; COB instead estimates multiscale contours and orientations in one image-level pass.The cited approaches include Ganin and Lempitsky,, Shen et al., and Xie and Tu.Source: §2 · p. 3
• COB uses both VGGNet and ResNet base architectures, supporting its modularity and potential to benefit from improvements to classification CNNs.The passage places these architectures among a progression from AlexNet and GoogLeNet to ResNets.Source: §2 · p. 3
• Other contour-learning work uses weak supervision from object detectors or video motion boundaries, whereas COB uses full supervision for contour localization and orientation.Khoreva et al. and Li et al. are the cited weakly supervised or unsupervised examples.Source: §2 · pp. 3–4
• Unlike prior hierarchy methods, COB combines region hierarchies from CNN contour outputs and bypasses normalized-cuts globalization, using CNN high-level information for strength and orientation.The comparison discusses Najman and Schmitt, Arbeláez et al., and Pont-Tuset et al..Source: §2 · p. 4

3 Deep Multiscale Oriented Contours

• The five-stage base CNN supplies feature maps at different scales, which COB combines to detect multiscale contours and orientations.Deeper features have lower resolution and larger receptive fields, providing more global information for predicting boundary strength and orientation.Source: §3 · pp. 4–5
• Each stage’s side activation is supervised as an intermediate contour map, and trainable weights combine activations into fine- and coarse-scale outputs.The fine output better localizes contours, while the coarse output produces less noisy detections; the base CNN weights are frozen while training the combination weights.Source: §3 · pp. 5–6
• For orientation estimation, COB adds K branches that classify contour pixels into orientation bins using feature maps from M scales.Ground-truth contour orientations are assigned from the nearest polygon segment, and the base-network weights remain frozen during branch training.Source: §3 · p. 6
• The orientation map combines responses from orientation bins, averaging neighboring-bin angles when both responses are strong and assigning random angles where none respond.Figure 3 illustrates four of the eight bin responses and the resulting orientation map.Source: §3 · p. 6, Figure 3 · p. 7
• Learned orientations are reported as significantly more accurate than local-gradient orientations and to lead to better region segmentations.The comparison is with orientations computed using local gradient filters.Source: §3 · p. 6

Fast Hierarchies from Multiscale Oriented Contours

• The sparse representation reduces multiscale UCM computation time by an order of magnitude, to less than one second.It also permits operations at all contour strengths, rather than discretizing UCM manipulation to a low number of strengths as in.Source: §4 · pp. 7–8

Sparse Boundary Representation of Hierarchies of Regions

• An Ultrametric Contour Map represents boundaries by the strength threshold at which each edgel disappears and its neighboring regions merge.Thresholding at different contour strengths yields partitions at different granularities.Source: §4 · pp. 7–8
• The sparse boundary representation stores neighboring-region pairs, boundary strengths, and boundary coordinates, enabling targeted operations without sweeping the full boundary grid.This avoids memory and runtime waste when only a small fraction of boundaries are active.Source: §4 · p. 8

Fast Hierarchies from Multiscale Oriented Contours

• Combining hierarchies from contour signals at each CNN layer, rather than combining contours directly, builds a single hierarchy from outputs computed in one CNN pass.The approach adapts the framework in, which combines UCMs from contours computed at different image scales.Source: §4 · p. 8
• Learned contour orientations are used to compute the Oriented Watershed Transform, further boosting performance.The method combines region hierarchies from layer-specific contour signals rather than combining the contour signals themselves.Source: §4 · p. 8

5.1 Control Experiments/Ablation Analysis

• ResNet50 improves boundaries and regions over VGGNet; with trained orientations it performs best, while removing globalization improves performance and speeds the system.Controls compare MCG with Structured Edges, MCG-HED, and direct five-side-output VGGNet and ResNet50 baselines.Source: §5.1 · pp. 9–10, Table 1 · p. 10

5.2 Contour Orientation

• 58.6% vs 41.2% and 12.5%: COB's orientation-accuracy AUC exceeds local gradients and random guessing.Accuracy is averaged across the eight orientation classes, which are imbalanced.Source: Figure 6 · p. 10, §5.2 · p. 10

5.3 Generic Image Segmentation

• On PASCAL Context, COB leads all compared methods for boundaries and regions; it also reaches state-of-the-art results on BSDS500.The paper reports smaller margins on BSDS500, where contour detection is near human performance.Source: Figure 7 · p. 11, §5.3 · p. 11, Figure 8 · p. 12

5.4 Object Proposals

• In PASCAL Segmentation and untrained MS-COCO, COB's average-recall improvement is at least +13% absolute over second-best methods.The comparisons span the full range of proposal counts.Source: §5.4 · p. 12, Figure 9 · p. 13

5.5 Efficiency Analysis

• COB’s full-system timing improves over previous state-of-the-art from 24.37 to 0.79 seconds; the final technique removes globalization, computes three scales in one pass, and adds contour orientations.Sparse boundaries reduce UCM and OWT time from 11.58 to 1.63 seconds. The network’s forward pass computes contour strength and eight orientations in 0.28 seconds on an NVidia Titan X GPU.Source: §5.5 · pp. 12–13

• The method detects multiscale contours and orientations in one CNN pass, constructs region hierarchies with a sparse boundary representation, and shifts evaluation from BSDS to PASCAL because BSDS performance is near saturation.Source: §6 · p. 14
• COB achieves state-of-the-art performance by a significant margin for contour detection, orientation estimation, generic image segmentation, and object proposals.Source: §6 · p. 14
• Using two CNN base architectures demonstrates COB's modularity and suggests future base-network improvements can transfer to perceptual grouping; it also avoids globalization, a speed bottleneck in previous approaches.Source: §6 · p. 14
• The authors make code, CNN models, precomputed results, dataset splits, and benchmarks publicly available.Source: §6 · p. 14