Research map / Label-efficient and robust learning / Recognition and long-tail
Coarse-to-Fine: Learning Compact Discriminative Representation for Single-Stage Image Retrieval
Yunquan Zhu, Xinkai Gao, Bo Ke, Ruizhi Qiao, Xing Sun
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 |
|---|---|---|---|
| replace ArcFace with MadaCos adaptive loss (ablation) | CFCD R50 · ROxford Hard | mAP: 58.53 -> 64.38 | Improved |
| add triplet loss with local descriptor matching after global training (ablation) | CFCD R50 · RParis Hard | mAP: 80.21 -> 81.68 | Improved |
| apply triplet loss from the start of training instead of from epoch 50 (ablation) | CFCD R50 · ROxford Hard | mAP: 64.84 -> 58.29 | Hurt |
| add hard negative sampling (ablation) | CFCD R50 · ROxford Hard | mAP: 64.84 -> 65.06 | Improved |
| add matching constraints to the triplet loss (ablation) | CFCD R50 · ROxford Hard | mAP: 62.37 -> 65.06 | Improved |
| use MadaCos instead of ArcFace in the coarse-to-fine framework (ablation) | CFCD R50 · ROxford Hard | mAP: 60.46 -> 65.06 | Improved |
| use dynamically tuned scale and margin instead of fixed converged values (ablation) | MadaCos · ROxford Hard | mAP: 50.28 -> 63.36 | Improved |
| use MadaCos instead of AdaCos (ablation) | global loss · ROxford Hard | mAP: 56.62 -> 63.36 | Improved |
| use a smaller matching threshold tau (ablation) | triplet matching · ROxford/RParis Medium | mAP: tau 30: 82.42 vs tau 70: 81.88 on ROxford Medium | Mixed |
| use fewer negatives per triplet (Q) (ablation) | triplet loss · ROxford Hard | mAP: 64.83 (Q=10) -> 66.03 (Q=2) | Improved |
| increase MadaCos anchor rho beyond 0.1 (ablation) | MadaCos · ROxford Hard | mAP: 61.25 (rho 0.1) -> 57.02 (rho 0.4) | Hurt |
| second-stage fine-tuning with longer total training (T=25 vs E=T=20) (ablation) | R50-CFCD 2048-d · ROxford Hard | mAP: 59.99 -> 63.59 | Improved |
Closest papers by meaning
- Towards Optimal Fine Grained Retrieval via Decorrelated Centralized Loss with Normalize-Scale Layer (AAAI 2019 · similarity 0.90)
- Unicom: Universal and Compact Representation Learning for Image Retrieval (ICLR 2023 · similarity 0.89)
- Grafit: Learning Fine-Grained Image Representations With Coarse Labels (ICCV 2021 · similarity 0.89)
- Instance-Level Composed Image Retrieval (NeurIPS 2025 · similarity 0.88)
- NUDGE: Lightweight Non-Parametric Fine-Tuning of Embeddings for Retrieval (ICLR 2025 · similarity 0.87)
- Exploiting Distribution Constraints for Scalable and Efficient Image Retrieval (ICLR 2025 · similarity 0.87)
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.
