Research map / Detection and segmentation / Open-set and few-shot detection
Visiting the Invisible: Layer-by-Layer Completed Scene Decomposition
Chuanxia Zheng, Duy-Son Dao, Guoxian Song, Tat‐Jen Cham, Jianfei Cai
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train decomposition and completion jointly instead of separately (ablation) | CSDNet (PICNet-patch, HTC) · CSD test | mask AP: 43.8 -> 50.3 | Improved |
| add scene completion between decomposition steps (ablation) | CSDNet (HTC) · CSD test | mask AP: 36.8 -> 50.3 | Improved |
| use patch attention instead of point attention in completion (ablation) | CSDNet (joint training) · CSD test | mask AP: 47.7 -> 50.3 | Improved |
| use HTC instead of Mask-RCNN as decomposition network (ablation) | CSDNet completion · CSD completion | PSNR: 27.31 -> 30.45 | Improved |
| train completion jointly with HTC decomposition instead of separately (ablation) | CSDNet completion · CSD completion | PSNR: 26.75 -> 30.45 | Improved |
| replace one-pass HTC with layer-by-layer CSDNet | HTC · CSD test | mask AP: 47.3 -> 50.3 | Improved |
| use ground-truth completed images (oracle) in each step | CSDNet (HTC) · CSD test | mask AP: 50.3 -> 56.0 | Improved |
| switch backbone from Mask-RCNN to HTC | CSDNet · CSD test | mask AP: 48.7 -> 50.3 | Improved |
| order by predicted layer order instead of mask area (ablation) | CSDNet · CSD test | OAP: 44.7 -> 81.7 | Improved |
| order by predicted layer order instead of y-axis position (ablation) | CSDNet · CSD test | OAP: 62.0 -> 81.7 | Improved |
| use CSDNet instead of MLC with HTC segmentation | amodal segmentation · COCOA | mAP: 34.4 -> 34.8 | Improved |
| use CSDNet instead of MLC with HTC segmentation | amodal segmentation · KINS | mAP: 31.6 -> 32.2 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SeGAN — SeGAN: Segmenting and Generating the Invisible (CVPR 2018) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Self-Supervised Scene De-Occlusion (CVPR 2020 · similarity 0.91)
- ObjectGS: Object-aware Scene Reconstruction and Scene Understanding via Gaussian Splatting (ICCV 2025 · similarity 0.89)
- Semantic Scene Completion via Integrating Instances and Scene In-the-Loop (CVPR 2021 · similarity 0.89)
- RevealNet: Seeing Behind Objects in RGB-D Scans (CVPR 2020 · similarity 0.89)
- VisHall3D: Monocular Semantic Scene Completion from Reconstructing the Visible Regions to Hallucinating the Invisible Regions (ICCV 2025 · similarity 0.88)
- Amodal Scene Analysis via Holistic Occlusion Relation Inference and Generative Mask Completion (AAAI 2024 · similarity 0.88)
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.
