Robust Object Detection under Occlusion with Context-Aware CompositionalNets

1. Introduction

• Faster R-CNN is unreliable under partial occlusion because proposals mislocalize objects and classification is not robust; CompositionalNets also lack context separation and robust box estimation.Context bias can amplify when detection thresholds are lowered for strongly occluded objects, producing false positives.Source: §1 · pp. 1–2
• The method decomposes image representation into context and object components, using bounding-box annotations to learn a context-aware representation that controls context influence.This is intended to improve robustness when classifying strongly occluded objects.Source: §1 · p. 2
• Part-based voting is extended to estimate two opposite bounding-box corners as well as the object center, enabling bounding-box estimation under severe occlusion.The proposed voting mechanism leverages unoccluded object parts.Source: §1 · p. 2
• 41% on PASCAL3D+ and 35% on MS-COCO in absolute performance relative to Faster R-CNN for strongly occluded vehicles.The proposed context-aware CompositionalNets use robust bounding-box estimation and detect objects even under severe occlusion.Source: §1 · p. 2

2. Related Work

• Region Proposal Networks are widely used for fast detection, but experiments show they do not estimate object bounding boxes correctly under occlusion.The paper identifies this as a limitation of region selection under occlusion.Source: §2 · p. 2
• Data augmentation with partial occlusion [8] [38] has limited impact on DCNN generalization, whereas CompositionalNets were shown to be more robust for image classification under occlusion.Earlier work also reports reduced robustness of standard DCNNs relative to humans [19].Source: §2 · p. 2
• Prior detection approaches include handcrafted-feature cascades, methods requiring part-level annotations [39], and 3D-based occlusion modeling.Other approaches use video or stereo, while part-based voting methods assume fixed-size bounding boxes.Source: §2 · pp. 2–3
• The proposed detector extends CompositionalNets with context awareness and part-based voting to estimate bounding boxes under very strong partial occlusion.The related-work discussion distinguishes this approach from methods assuming fixed-size bounding boxes.Source: §2 · p. 3

3.1. Prior work: CompositionalNets

• CompositionalNets replace a VGG-16 fully connected head with a generative compositional model whose part votes support robust classification under partial occlusion.The model represents feature activations as a mixture of compositional models.Source: §3.1 · p. 3

3.2. Detection with Robust Bounding Box Voting

• Scanning the feature map with the object model yields a spatial likelihood map, whose thresholded maxima after non-maximum suppression provide detections.The detection layer accumulates votes for the object center across feature-map positions.Source: §3.2 · p. 4
• Voting for the object center and two opposite bounding-box corners enables box estimation from visible parts despite substantial occlusion.The model generates likelihood maps for the center and corners, then uses the two points with maximal likelihood to form a box.Source: §3.2 · p. 4

3.3. Context-aware CompositionalNets

• Context-aware CompositionalNets represent context and object with separate models, allowing the context’s influence on detection to be controlled.Bounding-box annotations support training-image segmentation; features with receptive fields outside the boxes are treated as context.Source: §3.3 · p. 5
• The context-object trade-off is set by a prior ω at test time; ω = 0.5 retains the original CompositionalNet.The context and object parameters are learned from training data.Source: §3.3 · p. 5

3.4. Training Context-Aware CompositionalNets

• End-to-end training jointly optimizes generative fit, localization, and classification objectives using backpropagation.Context assignments are estimated beforehand from segmentation; the stated objectives are generative likelihood, detection, and classification.Source: §3.4 · pp. 5–6

4. Experiments

• The proposed vehicle-classification dataset contains 6 classes at fixed scale (224 pixels), with 9 occlusion levels spanning foreground object occlusion (20–80%) and background context occlusion (0–60%); OccludedCOCO contains 2,890 light- and 788 heavy-occlusion images.Evaluation uses only the highest-confidence bounding-box proposal and measures IoU at 50%.Source: §4 · pp. 6–7

4.1. Object Detection under Simulated Occlusion

• 59.9%: Faster R-CNN with CutOut still loses this much performance under strong occlusion versus the nonocclusion setup.The authors attribute failures to inaccurate RPN proposals and difficulty classifying valid object regions under heavy occlusion.Source: §4.1 · p. 7, Table 1 · p. 7
• With RPN proposals, CompositionalNet performs better than Faster R-CNN at high occlusion but worse at low occlusion.The authors suggest that RPN proposals are not precise enough for CompositionalNet, which requires close alignment to the object center.Source: §4.1 · pp. 7–8, Table 1 · p. 7
• Bounding-box voting substantially improves CompositionalNet performance over RPN proposals on OccludedVehiclesDetection.The result supports the finding that CompositionalNet classification requires precise proposals, including under partial occlusion.Source: Table 1 · p. 7, §4.1 · p. 8, Figure 7 · p. 8
• On the simulated dataset, performance decreases as background occlusion rises for both ω = 0.5 and ω = 0.The authors also report reduced detection precision at ω = 0.5.Source: Table 1 · p. 7, §4.1 · p. 8

4.2. Object Detection under Realistic Occlusion

• On OccludedCOCO, CA-CompositionalNet with robust bounding-box voting outperforms Faster R-CNN and CompNet+RPN.With ω = 0.2, context is helpful under strong occlusion but slightly decreases performance under low occlusion.Source: Table 2 · p. 7, §4.2 · p. 8, Figure 8 · p. 8

• Standard proposal networks are increasingly misguided as occluders occupy more context, while classification networks also fail to classify partially occluded objects robustly.Source: §5 · p. 8
• A detection layer generalizes CompositionalNets, previously shown to classify partially occluded objects robustly, to object detection.Source: §5 · p. 8
• Part-based voting uses unoccluded object parts to estimate bounding boxes accurately, even under severe occlusion.Source: §5 · p. 8
• Bounding-box annotations support object-context segmentation and end-to-end learning of a representation that disentangles context from the object.Source: §5 · p. 8