Source-linked AI summary
Deep Occlusion-Aware Instance Segmentation with Overlapping BiLayers
Lei Ke, Yu-Wing Tai, Chi-Keung Tang
TL;DR
Highly overlapping objects confuse real contours with occlusion boundaries, limiting conventional mask regression. BCNet models occluders and occludees in interacting bilayer graph layers, and experiments report consistent gains across segmentation settings, especially under heavy occlusion.
Problem
Highly overlapping objects create mask conflicts because conventional approaches do not distinguish real object contours from occlusion boundaries or explicitly model occluders.
Method
BCNet uses cascaded bilayer GCN layers that jointly model occluder and occludee masks and boundaries within the same ROI.
Results
BCNet achieves consistent gains across modal and amodal instance segmentation with different backbones and detectors, including heavy-occlusion cases.
Takeaways & Limitations
Explicit occluder–occludee modeling decouples overlapping instances into separate graph spaces while preserving their interaction during segmentation.
Abstract
from arXiv · showhide
Segmenting highly-overlapping objects is challenging, because typically no distinction is made between real object contours and occlusion boundaries. Unlike previous two-stage instance segmentation methods, we model image formation as composition of two overlapping layers, and propose Bilayer Convolutional Network (BCNet), where the top GCN layer detects the occluding objects (occluder) and the bottom GCN layer infers partially occluded instance (occludee). The explicit modeling of occlusion relationship with bilayer structure naturally decouples the boundaries of both the occluding and occluded instances, and considers the interaction between them during mask regression. We validate the efficacy of bilayer decoupling on both one-stage and two-stage object detectors with different backbones and network layer choices. Despite its simplicity, extensive experiments on COCO and KINS show that our occlusion-aware BCNet achieves large and consistent performance gain especially for heavy occlusion cases. Code is available at https://github.com/lkeab/BCNet.
1. Introduction
Highly overlapping objects create mask conflicts because standard instance segmentation often treats occlusion boundaries as object contours. BCNet addresses this by jointly modeling occluders and occludees in separate interacting layers.
- Motivation: Overlapping same-class objects cause substantial segmentation errors when their contours and occlusion boundaries are difficult to distinguish.Traditional mask heads can include occluding regions as part of the occluded object, leaving mask conflicts unresolved.
- BCNet: BCNet simultaneously regresses occluding regions and partially occluded objects after ROI extraction in two separate image layers.The top layer handles the occluder, while the bottom layer handles the occludee.
- BCNet: The bilayer structure decouples object boundaries and models occluder–occludee interaction during mask regression.
- Motivation: Previous methods use suppression or post-processing to resolve neighboring-mask conflicts, producing over-smooth boundaries or small gaps.
- Evaluation: The evaluation includes a COCO subset containing occluding and partially occluded objects, alongside original COCO evaluation.The paper also introduces large-scale occlusion-aware datasets with ground-truth complete contours for both object types.
- Results: BCNet outperforms state-of-the-art methods on both modal and amodal instance segmentation tasks.
2. Related Work
Prior work spans two-stage, one-stage, occlusion-handling, and amodal segmentation methods, but BCNet explicitly represents occluder–occludee relationships within a bilayer mask head.
- Instance Segmentation: Two-stage methods detect bounding boxes first and then segment each instance within its ROI, achieving state-of-the-art performance.
- Instance Segmentation: One-stage methods remove box detection and feature re-pooling, improving efficiency but tending to be less accurate than two-stage methods.
- Occlusion Handling: Existing occlusion-handling methods model spatial constraints, deformable templates, 3D shape, or occlusion overlap scores across application settings.
- Occlusion Handling: BCNet explicitly models occlusion patterns in shape and appearance to provide occlusion perception and reasoning.
- BCNet: The bilayer approach integrates into state-of-the-art segmentation frameworks for end-to-end training.
- Amodal Instance Segmentation: Unlike methods that directly regress a single occluded boundary, BCNet predicts complete segments in two graph layers and guides the occludee with occluder shape and location.
3. Occlusion-Aware Instance Segmentation
BCNet addresses heavy-occlusion instance segmentation by explicitly modeling occluder and occludee regions as interacting bilayer graph representations. Its cascaded GCN mask head separates occlusion boundaries from object contours and is trained jointly with detection and segmentation objectives.
- Motivation: Heavy occlusion confuses real object contours with occlusion boundaries, while conventional mask heads neglect occluding instances and overlap relations.The framework targets mask conflicts arising when multiple same-class objects share an ROI.
- Framework and Learning: The complete framework combines FPN-based ROI features, an FCOS detection head, and BCNet in an end-to-end multi-task optimization.The objective includes detection, occluder boundary and mask, and occludee boundary and mask losses, balanced by tuned hyperparameters.
- Bilayer GCN Structure: Graph convolution propagates information globally across ROI pixels using feature-similarity adjacency and residual nonlinear transformations.Each graph node represents a pixel, allowing separated portions of an occluded object to exchange information across the occluder.
- Bilayer GCN Structure: BCNet uses two orthogonal graph layers to represent occluder and occludee regions within the same ROI.The bilayer structure gives overlapping pixels two concurrent semantic states rather than only foreground/background labels.
- Occluder-Occludee Modeling: The first GCN predicts occluder contours and masks, then passes occlusion-aware features to the second GCN for occludee mask and boundary prediction.The cascaded design uses occluder shape and position information to guide partially occluded object segmentation.
- Framework and Learning: Training balances occlusion cases by filtering non-occluded ROI proposals so occlusions comprise 50% of sampled cases.The reported training setup uses SGD with momentum for 90K iterations, with 1K warm-up iterations and batch size 16.
4. Experiments
Experiments evaluate BCNet on COCO, COCO-OCC, COCOA, and KINS, including ablations of explicit occlusion modeling, bilayer design, GCN choice, detectors, and qualitative behavior. Results show consistent gains, particularly for heavily overlapping objects and amodal segmentation.
- 4.1. Experimental Setup: COCO training uses 2017train, while COCO-OCC contains 1,005 validation images selected for bounding-box overlap of at least 0.2.COCO results use standard metrics on validation and test-dev; COCO-OCC targets highly overlapping objects.
- 4.1. Experimental Setup: The synthetic occlusion dataset contains 100k images with complete contours for both occluders and partially occluded objects.Its annotations explicitly represent the occlusion relationship between occluding regions and occluded objects.
- 4.2. Ablation Study: Joint occlusion modeling raises mask AP from 32.65 to 33.43 on standard COCO validation and from 29.04 to 30.37 on COCO-OCC.The comparison is against a BCNet baseline without explicit occlusion-modeling targets, with the largest improvement reported for heavy occlusion.
- 4.2. Ablation Study: On COCO-OCC, bilayer GCN improves AP from 29.63 to 30.68 over single GCN, while bilayer FCN improves single FCN from 28.43 to 30.12.A second GCN layer uses occlusion-aware feature guidance for final occludee segmentation, adding 1.23 AP on COCO-OCC and 1.06 AP on COCO.
- 4.3. Performance Comparison and Analysis: BCNet outperforms state-of-the-art methods on modal and amodal segmentation, including 31.71 AP versus 30.32 for Mask Scoring R-CNN on COCO-OCC with Faster R-CNN.Training on the synthetic occlusion dataset further increases COCO-OCC AP to 32.89 and AP50 to 53.25; qualitative results show improved boundary estimation and recovered missed parts.
5. Conclusion
BCNet addresses highly overlapping-instance segmentation by explicitly modeling occluder–occludee relationships, producing consistent gains across detector settings and segmentation modes.
- BCNet achieves consistent overall segmentation gains across different backbones and object detectors in both modal and amodal settings.
- Explicit occluder–occludee modeling decouples occluding and occluded instances into two disjoint graph spaces.
- The bilayer structure explicitly considers interactions between objects within each ROI region during segmentation.
- The authors identify BCNet as beneficial for future research in occlusion handling and instance segmentation.