Research map / Detection and segmentation / Weakly supervised detection and segmentation
Learning to Segment via Cut-and-Paste
Tal Remez, Jonathan Huang, Matthew Brown
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 |
|---|---|---|---|
| paste along the same horizontal scanline instead of uniformly in x and y (ablation) | Cut&Paste · Cityscapes | per-pixel mask accuracy: 68% to 72% | Improved |
| paste Cityscapes objects into random COCO images (ablation) | Cut&Paste · Cityscapes | per-pixel mask accuracy: 60% | Hurt |
| composite at 2x mask resolution instead of mask resolution (ablation) | Cut&Paste · Cityscapes | per-pixel mask accuracy: 71% to 66% | Hurt |
| increase aerial training instances from 5K to 1M (ablation) | Cut&Paste · Aerial building segmentation | mask prediction error: about 10% | Improved |
| add classification loss weight w_cls in [1,4] vs 0 (ablation) | Cut&Paste · Aerial building segmentation | mask accuracy / stability: w_cls=0 poor and unstable; w_cls in [1,4] optimal | Improved |
| add classification loss weight on Cityscapes (ablation) | Cut&Paste · Cityscapes | mask accuracy / stability: increases stability but does not improve performance | No clear effect |
| vary cut loss weight w_cut (ablation) | Cut&Paste · Aerial building segmentation | mask accuracy: did not work as well as classification loss | Hurt |
| use Cut&Paste instead of Simple Does It on baseball bat | Cut&Paste · COCO | mIoU: 0.43 vs 0.32 | Improved |
| use Cut&Paste instead of Simple Does It on skis | Cut&Paste · COCO | mIoU: 0.27 vs 0.23 | Improved |
| use Cut&Paste instead of Simple Does It on kite | Cut&Paste · COCO | mIoU: 0.51 vs 0.56 | Hurt |
| use Cut&Paste instead of Simple Does It on airplane | Cut&Paste · COCO | mIoU: 0.48 vs 0.55 | Hurt |
| use Cut&Paste instead of GrabCut/Box on all COCO classes | Cut&Paste · COCO (all 80 classes) | mIoU: 0.61 -> 0.64 | Improved |
Closest papers by meaning
- Prior to Segment: Foreground Cues for Weakly Annotated Classes in Partially Supervised Instance Segmentation (ICCV 2021 · similarity 0.90)
- Pointly-Supervised Instance Segmentation (CVPR 2022 · similarity 0.90)
- SIM: Semantic-Aware Instance Mask Generation for Box-Supervised Instance Segmentation (CVPR 2023 · similarity 0.90)
- Learning to Segment Every Thing (CVPR 2018 · similarity 0.89)
- BoxTeacher: Exploring High-Quality Pseudo Labels for Weakly Supervised Instance Segmentation (CVPR 2023 · similarity 0.89)
- Boundary-Aware Instance Segmentation (CVPR 2017 · similarity 0.89)
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.
