Research map / Detection and segmentation / Open-set and few-shot detection
BLADE: Box-Level Supervised Amodal Segmentation through Directed Expansion
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes BLADE, a box-level supervised amodal segmentation network that expands visible masks into amodal masks using an overlapping-region-based hybrid of three branches and a connectivity loss.
Problem
Amodal segmentation requires pixel-level amodal annotations that are costly; existing box-level weakly supervised methods produce low-resolution masks with imprecise boundaries.
Approach
Decompose amodal mask into visible and occluded parts; use the overlapping region between instance amodal boxes to route predictions (coarse amodal mask inside overlapping region, visible mask elsewhere), and add a connectivity loss (neighbor + uniform terms) that directs expansion of visible masks into occluded areas.
Key results
- On OccludedVehicles, BLADE reaches 68.2 mean IoU vs 57.6 for Bayesian-Amodal (known c), and over 12% better than unknown-c Bayesian-Amodal.
- On KINS, BLADE achieves 73.4 mean IoU under weak supervision, improving over Bayesian-Amodal (unknown c, 62.1) by over 11%.
- On COCOA-cls transfer, BLADE reaches 71.2 mean IoU vs 57.6 for Bayesian-Amodal (unknown c).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use BLADE instead of Bayesian-Amodal (unknown c) | Box-level supervised amodal segmentation · OccludedVehicles | mean IoU: 56.0 -> 68.2 | Helped |
| Use BLADE instead of Bayesian-Amodal (unknown c) | Box-level supervised amodal segmentation · KINS | mean IoU: 62.1 -> 73.4 | Helped |
| Use BLADE instead of Bayesian-Amodal (unknown c) | Box-level supervised amodal segmentation · COCOA-cls | mean IoU: 57.6 -> 71.2 | Helped |
| Add the fusion structure and region branch (ablation) | BLADE · KINS | mean IoU: 65.9 -> 73.4 | Helped |
| Add the uniform loss (ablation) | BLADE · KINS | mean IoU: 71.0 -> 73.4 | Helped |
| Add the neighbor loss (ablation) | BLADE · KINS | mean IoU: 63.6 -> 73.4 | Helped |
| Add the uniform and neighbor losses (ablation) | BLADE without fusion · KINS | mean IoU: 62.9 -> 65.9 | Helped |
| Set Lᵃ weights to (2.0, 1.0, 1.0) rather than (1.0, 1.0, 1.0) (ablation) | BLADE · KINS | mean IoU: 71.3 -> 73.4 | Helped |
| Set Lᵃ weights to (1.0, 2.0, 1.0) rather than (2.0, 1.0, 1.0) (ablation) | BLADE · KINS | mean IoU: 73.4 -> 73.2 | No clear effect |
| Set Lᵃ weights to (1.0, 1.0, 2.0) rather than (2.0, 1.0, 1.0) (ablation) | BLADE · KINS | mean IoU: 73.4 -> 70.0 | Hurt |
Limitations
- Weakly supervised method still lags fully supervised methods (e.g., AISFormer 76.7 mean IoU on KINS vs 73.4).
- Relies on overlapping-region envelope boxes derived from amodal boxes and on predicted visible masks as cues for expansion.
- Evaluated on relatively small COCOA-cls (766 objects) and synthetic OccludedVehicles datasets.
Experimental setup
Datasets: OccludedVehicles KINS COCOA-cls
Benchmarks: OccludedVehicles KINS COCOA-cls
Metrics: mean IoU
Baselines: BBTP BoxInst Bayesian-Amodal SAM VRSP AISFormer
Builds on: BoxInst Bayesian-Amodal FCOS Detectron2
Closest papers by meaning
- Amodal Ground Truth and Completion in the Wild (CVPR 2024 · similarity 0.91)
- Segment Anything, Even Occluded (CVPR 2025 · similarity 0.89)
- Towards Efficient Foundation Model for Zero-shot Amodal Segmentation (CVPR 2025 · similarity 0.88)
- Using Diffusion Priors for Video Amodal Segmentation (CVPR 2025 · similarity 0.88)
- pix2gestalt: Amodal Segmentation by Synthesizing Wholes (CVPR 2024 · similarity 0.88)
- Amodal Scene Analysis via Holistic Occlusion Relation Inference and Generative Mask Completion (AAAI 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.