Research map / Detection and segmentation / Weakly supervised detection and segmentation
BoxInst: High-Performance Instance Segmentation With Box Annotations
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Replace box-generated masks with the projection loss (ablation) | BoxInst · COCO val2017 | mask AP: 10.6 -> 21.2 | Helped |
| Add the pairwise affinity loss to the projection loss (ablation) | BoxInst · COCO val2017 | mask AP: 21.2 -> 30.7 | Helped |
| Set color similarity threshold τ to 0 instead of 0.1 (ablation) | BoxInst pairwise loss · COCO val2017 | mask AP: 30.7 -> 9.4 | Hurt |
| Increase color similarity threshold τ from 0.1 to 0.2 (ablation) | BoxInst pairwise loss · COCO val2017 | mask AP: 30.7 -> 30.6 | Hurt |
| Use a 5 × 5 neighborhood instead of 3 × 3, both with dilation 1 (ablation) | BoxInst pairwise loss · COCO val2017 | mask AP: 29.7 -> 30.5 | Helped |
| Use dilation 2 instead of dilation 1 for a 3 × 3 neighborhood (ablation) | BoxInst pairwise loss · COCO val2017 | mask AP: 29.7 -> 30.7 | Helped |
| Apply dilation 2 to a 5 × 5 neighborhood instead of dilation 1 (ablation) | BoxInst pairwise loss · COCO val2017 | mask AP: 30.5 -> 29.9 | Hurt |
| Replace the original pixelwise Dice loss with projection and pairwise affinity losses (ablation) | CondInst, fully supervised · COCO val2017 | mask AP: 35.6 -> 35.4 | No clear effect |
| Use projection and pairwise losses for box-only unseen classes while training with masks for 20 classes (ablation) | BoxInst semi-supervised · COCO val2017, 60 unseen classes | mask AP: 19.9 -> 29.7 -> 30.9 | Helped |
| Add partial mask annotations for 20 classes to box-supervised training (ablation) | BoxInst · COCO val2017, 60 unseen classes | mask AP: 29.6 -> 30.9 | Helped |
Closest papers by meaning
- BoxTeacher: Exploring High-Quality Pseudo Labels for Weakly Supervised Instance Segmentation (CVPR 2023 · similarity 0.93)
- SIM: Semantic-Aware Instance Mask Generation for Box-Supervised Instance Segmentation (CVPR 2023 · similarity 0.92)
- Learning to Segment Every Thing (CVPR 2018 · similarity 0.92)
- Extreme Point Supervised Instance Segmentation (CVPR 2024 · similarity 0.90)
- Training-Free Dataset Pruning for Instance Segmentation (ICLR 2025 · similarity 0.89)
- Sketchy Bounding-box Supervision for 3D Instance Segmentation (CVPR 2025 · 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.