Research map / Detection and segmentation / Weakly supervised detection and segmentation
AttentionShift: Iteratively Estimated Part-Based Attention Map for Pointly Supervised Instance Segmentation
Read the paperCheck 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 |
|---|---|---|---|
| Use part-based attention maps with key-point shift and filtering instead of the BESTIE point-supervised approach | Pointly supervised instance segmentation · PASCAL VOC 2012 val | mAP@0.5: 46.7% vs 54.4% | Helped |
| Use AttnShift-S instead of BESTIE | Pointly supervised instance segmentation · MS COCO 2017 val | AP@0.5: 34.0% vs 38.8% | Helped |
| Replace the self-attention map with an instance attention map (ablation) | AttentionShift baseline · PASCAL VOC 2012 val | mAP: 38.0% -> 46.8% | Helped |
| Add key-point shift to decompose the attention map into part-based maps (ablation) | AttentionShift · PASCAL VOC 2012 val | mAP: 46.8% -> 49.2% | Helped |
| Add key-point filtering (ablation) | AttentionShift with key-point shift · PASCAL VOC 2012 val | mAP: 49.2% -> 53.2% | Helped |
| Set key-point shift temperature α to 0.1 instead of 1.0 (ablation) | AttentionShift · PASCAL VOC 2012 val | mAP: 53.2 -> 51.4 | Hurt |
| Set key-point shift temperature α to 10 instead of 1.0 (ablation) | AttentionShift · PASCAL VOC 2012 val | mAP: 53.2 -> 49.1 | Hurt |
| Use α=Inf, reducing the weighted shift to vanilla key-point shift (ablation) | AttentionShift · PASCAL VOC 2012 val | mAP: 53.2 -> 49.8 | Hurt |
| Use 0 shift iterations instead of 10 (ablation) | AttentionShift · PASCAL VOC 2012 val | mAP: 53.2 -> 50.3 | Hurt |
| Use 5 shift iterations instead of 10 (ablation) | AttentionShift · PASCAL VOC 2012 val | mAP: 53.2 -> 51.8 | Hurt |
| Use 20 shift iterations instead of 10 (ablation) | AttentionShift · PASCAL VOC 2012 val | mAP: 53.2 -> 52.5 | Hurt |
| Set the key-point filtering foreground threshold Tf to 0.9 (ablation) | AttentionShift · PASCAL VOC 2012 val | mAP: 53.2 -> 52.3 | Hurt |
Closest papers by meaning
- P2Seg: Pointly-supervised Segmentation via Mutual Distillation (ICLR 2024 · similarity 0.87)
- Repurposing Stable Diffusion Attention for Training-Free Unsupervised Interactive Segmentation (CVPR 2025 · similarity 0.86)
- Weakly Supervised Instance Segmentation Using Class Peak Response (CVPR 2018 · similarity 0.85)
- Semantic-aware SAM for Point-Prompted Instance Segmentation (CVPR 2024 · similarity 0.85)
- WISH: Weakly Supervised Instance Segmentation using Heterogeneous Labels (CVPR 2025 · similarity 0.84)
- Extreme Point Supervised Instance Segmentation (CVPR 2024 · similarity 0.84)
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.