Research map / Label-efficient and robust learning / Self-supervised and contrastive
Contrastive Attention Maps for Self-Supervised Co-Localization
Minsong Ki, Youngjung Uh, Junsuk Choe, Hyeran Byun
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 |
|---|---|---|---|
| add classification loss Lcls to CAML+PTAP (ablation) | Ours · CUB-200-2011, VGG16, rotation | CorLoc Mean: 70.74 -> 73.20 | Improved |
| add contrastive attention map loss Lcaml (ablation) | Baseline + PTAP · CUB-200-2011, VGG16, rotation | CorLoc Mean: 69.44 -> 73.20 | Improved |
| add pixel-wise top-k attention pooling (ablation) | CAML model · CUB-200-2011, VGG16, rotation | CorLoc Mean: 69.50 -> 73.20 | Improved |
| compute contrastive loss on attention maps instead of feature maps (ablation) | Lcaml · CUB-200-2011, VGG16, rotation | CorLoc Mean: 67.92 -> 73.20 | Improved |
| add negative pairs (ablation) | Lcaml · CUB-200-2011, VGG16, rotation | CorLoc Mean: 70.16 -> 73.20 | Improved |
| add positive pairs (ablation) | Lcaml · CUB-200-2011, VGG16, rotation | CorLoc Mean: 67.92 -> 73.20 | Improved |
| use top-70% pooling instead of max pooling (ablation) | PTAP · CUB-200-2011, VGG16, rotation | CorLoc Mean: 58.01 -> 73.20 | Improved |
| use top-70% pooling instead of average pooling (ablation) | PTAP · CUB-200-2011, VGG16, rotation | CorLoc Mean: 70.94 -> 73.20 | Improved |
| use top-70% pooling instead of bottom-30% pooling (ablation) | PTAP · CUB-200-2011, VGG16, rotation | CorLoc Mean: 54.89 -> 73.20 | Improved |
| add CAML+PTAP to rotation pretext baseline (ablation) | classification-only baseline · CUB-200-2011, VGG16, rotation | CorLoc IoU=0.5: 77.21 -> 83.65 (+6.4) | Improved |
| add CAML+PTAP to scale pretext baseline (ablation) | classification-only baseline · CUB-200-2011, VGG16, scale | CorLoc IoU=0.5: 44.66 -> 83.62 (+39.0) | Improved |
| add CAML+PTAP to translation pretext baseline (ablation) | classification-only baseline · CUB-200-2011, VGG16, translation | CorLoc IoU=0.5: 22.98 -> 83.05 (+60.1) | Improved |
Closest papers by meaning
- Self-Supervised Object Localization with Joint Graph Partition (AAAI 2022 · similarity 0.88)
- Attending on Attention ($A^2$): Smaller Self-Supervised ViTs Localize Better Than Larger Ones (NeurIPS 2026 · similarity 0.88)
- Unsupervised Object Localization with Representer Point Selection (ICCV 2023 · similarity 0.87)
- FOCUS: Forcing In-Context Object Localization through Visual Support Constraints and Policy Optimization (ICML 2026 · similarity 0.87)
- Co-Attention for Conditioned Image Matching (CVPR 2021 · similarity 0.87)
- SdalsNet: Self-Distilled Attention Localization and Shift Network for Unsupervised Camouflaged Object Detection (AAAI 2025 · similarity 0.86)
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.
