Research map / Label-efficient and robust learning / Self-supervised and contrastive
Learning Common Rationale To Improve Self-Supervised Representation for Fine-Grained Visual Recognition Problems
Yangyang Shu, Anton van den Hengel, Lingqiao Liu
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 a limited-capacity GradCAM fitting branch and use its attention mask for test-time weighted pooling | MoCo v2 · CUB-200-2011 | rank-1 retrieval: 42.72 -> 49.69 | Improved |
| add a limited-capacity GradCAM fitting branch and use its attention mask for test-time weighted pooling | MoCo v2 · Stanford Cars | Top-1 accuracy: 58.43 -> 60.75 | Improved |
| add a limited-capacity GradCAM fitting branch and use its attention mask for test-time weighted pooling | MoCo v2 · FGVC Aircraft | Top-1 accuracy: 52.54 -> 55.87 | Improved |
| omit weighted average pooling at test time, retaining GradCAM fitting as an auxiliary task (ablation) | Ours · CUB-200-2011 | rank-1 retrieval: 49.69 -> 42.93 | Hurt |
| apply weighted average pooling during both training and testing (ablation) | Ours · fine-grained datasets | retrieval performance | Hurt |
| replace the max-projection GFB with a higher-capacity MLP (ablation) | Ours · Stanford Cars | rank-1 retrieval: 34.56 -> 25.57 | Hurt |
| increase the number of projections to a very large value, beyond the reasonable range (ablation) | Ours · CUB-200-2011 | rank-1 retrieval | Hurt |
| set the KL loss weight ν to 1 instead of 0.01 (ablation) | Ours · CUB-200-2011 | rank-1 retrieval: 49.69 -> 47.52 | Hurt |
| set the KL loss weight ν to 0.1 instead of 0.01 (ablation) | Ours · CUB-200-2011 | rank-1 retrieval: 49.69 -> 48.83 | Hurt |
| set the KL loss weight ν to 0.001 instead of 0.01 (ablation) | Ours · CUB-200-2011 | rank-1 retrieval: 49.69 -> 46.63 | Hurt |
| combine the method with ContrastiveCrop | Ours · Stanford Cars | Top-1 accuracy: 60.75 -> 63.71 | Improved |
| combine the method with ContrastiveCrop | Ours · CUB-200-2011 | rank-1 retrieval: 49.69 -> 49.36 | No clear effect |
Closest papers by meaning
- Revisiting Self-Supervised Visual Representation Learning (CVPR 2019 · similarity 0.87)
- On Learning Discriminative Features from Synthesized Data for Self-Supervised Fine-Grained Visual Recognition (ECCV 2024 · similarity 0.86)
- Improving Fine-Grained Visual Recognition in Low Data Regimes via Self-Boosting Attention Mechanism (ECCV 2022 · similarity 0.86)
- Scaling and Benchmarking Self-Supervised Visual Representation Learning (ICCV 2019 · similarity 0.85)
- Collapse-Proof Non-Contrastive Self-Supervised Learning (ICML 2025 · similarity 0.83)
- When the Future Becomes the Past: Taming Temporal Correspondence for Self-supervised Video Representation Learning (CVPR 2025 · similarity 0.83)
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.
