Research map / Label-efficient and robust learning / Recognition and long-tail
Contextual Similarity Distillation for Asymmetric Image Retrieval
Check 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 |
|---|---|---|---|
| Distill contextual similarities with CSD instead of the strongest listed baseline, HVS | EfficientNetB3 query model with R101-DELG gallery · ROxf Medium | mAP: 76.41 -> 77.44 | Helped |
| Distill contextual similarities with CSD instead of the strongest listed baseline, HVS | EfficientNetB3 query model with R101-DELG gallery · ROxf Hard + R1M | mAP: 36.86 -> 38.20 | Helped |
| Distill contextual similarities with CSD instead of the strongest listed baseline, HVS | EfficientNetB3 query model with R101-DELG gallery · RPar Medium + R1M | mAP: 71.54 -> 73.37 | Helped |
| Distill contextual similarities with CSD instead of the strongest listed baseline, HVS | EfficientNetB3 query model with R101-DELG gallery · RPar Hard | mAP: 74.53 -> 75.68 | Helped |
| Use KL divergence rather than L1 contextual-similarity regression (ablation) | CSD with MobileNetV2 query and R101-GeM gallery · ROxf Medium | mAP: 60.7 -> 64.1 | Helped |
| Use KL divergence rather than L2 contextual-similarity regression (ablation) | CSD with MobileNetV2 query and R101-GeM gallery · RPar Hard | mAP: 53.6 -> 54.2 | Helped |
| Add second-order ranking-list consistency to first-order feature regression (ablation) | CSD with MobileNetV2 query and R101-GeM gallery · ROxf Medium | mAP: 50.3 -> 62.5 | Helped |
| Add first-order consistency to second-order KL consistency (ablation) | CSD with MobileNetV2 query and R101-GeM gallery · ROxf Medium | mAP: 60.4 -> 64.1 | Helped |
| Use near neighbors as anchors instead of randomly selected gallery images (ablation) | CSD with MobileNetV2 query and R101-GeM gallery · Oxford and Paris Medium | mAP | Helped |
| Increase the contextual-similarity ranking-list length (ablation) | CSD with MobileNetV2 query and R101-GeM gallery · ROxf and RPar | mAP | Helped |
| Add 10% GLDv2 training data to SfM120k (ablation) | CSD with MobileNetV2 query and R101-DELG gallery · ROxf Hard | mAP: 53.8 -> 55.9 | Helped |
| Increase GLDv2 training data from 10% to 40% (ablation) | CSD with MobileNetV2 query and R101-DELG gallery · ROxf Medium | mAP: 70.6 -> 75.2 | Helped |
Closest papers by meaning
- D3still: Decoupled Differential Distillation for Asymmetric Image Retrieval (CVPR 2024 · similarity 0.88)
- Exploiting Distribution Constraints for Scalable and Efficient Image Retrieval (ICLR 2025 · similarity 0.86)
- QuARI: Query Adaptive Retrieval Improvement (NeurIPS 2025 · similarity 0.86)
- LOCORE: Image Re-ranking with Long-Context Sequence Modeling (CVPR 2025 · similarity 0.85)
- Let All Be Whitened: Multi-Teacher Distillation for Efficient Visual Retrieval (AAAI 2024 · similarity 0.85)
- Image2Sentence based Asymmetrical Zero-shot Composed Image Retrieval (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.