Research map / Label-efficient and robust learning / Recognition and long-tail
Learning With Average Precision: Training Image Retrieval With a Listwise Loss
Jerome Revaud, Jon Almazan, Rafael S. Rezende, Cesar Roberto de Souza
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 triplet loss with hard negative mining by the listwise mAP loss (ablation) | GeM ResNet-101 · ROxford Medium | mAP: 64.9 -> 67.4 | Improved |
| replace triplet loss with hard negative mining by the listwise mAP loss (ablation) | GeM ResNet-101 · RParis Hard | mAP: 58.7 -> 61.0 | Improved |
| increase triplet batch from 64 to 1024 triplets (ablation) | GeM (TL) · ROxford Medium | mAP: 64.9 -> 65.5 | No clear effect |
| add class weighting inside each batch (ablation) | GeM (AP) · Revisited Oxford/Paris Medium | mAP: +2% | Improved |
| use tie-aware AP formulation instead of plain quantized AP (ablation) | GeM (AP) · Revisited Oxford/Paris Medium | mAP | Hurt |
| increase batch size up to 4096 (ablation) | GeM (AP) · Revisited Oxford/Paris Medium | mAP: saturates beyond 4096 | Improved |
| sample batch from a limited random set of classes (ablation) | GeM (AP) · Revisited Oxford/Paris Medium | mAP | No clear effect |
| vary number of quantization bins M (ablation) | GeM (AP) · Revisited Oxford/Paris Medium | mAP: M=20 used | No clear effect |
| add multi-scale testing (ablation) | GeM (AP) · Revisited Oxford/Paris | mAP | No clear effect |
| add weighted query expansion | GeM (AP) · ROxford Medium | mAP: 67.5 -> 71.4 | Improved |
| fine-tune with mAP loss instead of off-the-shelf features | GeM ResNet-101 · ROxford Medium | mAP: 45.0 -> 67.5 | Improved |
| use listwise mAP loss instead of triplet loss for training cost (ablation) | GeM ResNet-101 · Training cost | training time: 3 days -> 1 day | Improved |
Closest papers by meaning
- Optimization of Rank Losses for Image Retrieval (TPAMI 2025 · similarity 0.90)
- Robust and Decomposable Average Precision for Image Retrieval (NeurIPS 2021 · similarity 0.90)
- Hierarchical Average Precision Training for Pertinent Image Retrieval (ECCV 2022 · similarity 0.90)
- Smooth-AP: Smoothing the Path Towards Large-Scale Image Retrieval (ECCV 2020 · similarity 0.87)
- Rethinking the Optimization of Average Precision: Only Penalizing Negative Instances before Positive Ones Is Enough (AAAI 2022 · similarity 0.87)
- Threshold-Consistent Margin Loss for Open-World Deep Metric Learning (ICLR 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.
