Research map / Vision-language and multimodal / Human and embodied modeling
Learning Token-Based Representation for Image Retrieval
Read the paperCheck 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 |
|---|---|---|---|
| Use token-based local feature aggregation with refinement | R101-Ours · ROxf, ROxf+R1M, RPar, and RPar+R1M, Medium and Hard | mAP: 82.28 vs. 81.78; 70.52 vs. 70.17; 89.34 vs. 88.63; 76.66 vs. 76.04; 66.57 vs. 64.77; 55.90 vs. 53.69; 78.56 vs. 76.80; 5 | Helped |
| Add a tokenizer to the spatial-pooling baseline (ablation) | R101-SPoC baseline · ROxf Medium | mAP: 77.0 -> 79.8 | Helped |
| Add LFSA to the tokenizer model (ablation) | Tokenizer model · ROxf Medium | mAP: 79.8 -> 80.4 | Helped |
| Add refinement to the tokenizer model (ablation) | Tokenizer model · ROxf Medium | mAP: 79.8 -> 81.3 | Helped |
| Add tokenizer, LFSA, and refinement (ablation) | R101-SPoC baseline · ROxf Medium | mAP: 77.0 -> 82.3 | Helped |
| Add self-attention and cross-attention in the refinement block (ablation) | Tokenized model · ROxf Medium | mAP: 80.4 -> 82.3 | Helped |
| Add self-attention while retaining cross-attention (ablation) | Tokenized model · ROxf Medium | mAP: 80.9 -> 82.3 | Helped |
| Add cross-attention while retaining self-attention (ablation) | Tokenized model · ROxf Medium | mAP: 81.3 -> 82.3 | Helped |
| Replace attention tokenizer with VQ-based tokens (ablation) | Tokenized model · ROxf Medium | mAP: 82.3 -> 79.4 | Hurt |
| Replace attention tokenizer with learned tokens (ablation) | Tokenized model · ROxf Medium | mAP: 82.3 -> 81.1 | Hurt |
| Set the token count to 1, 2, or 3 instead of 4 (ablation) | Attention-tokenized model · ROxf Medium | mAP: 82.3 -> 79.8; 82.3 -> 80.3; 82.3 -> 81.6 | Hurt |
| Set the token count to 6 or 8 instead of 4 (ablation) | Attention-tokenized model · ROxf Medium | mAP: 82.3 -> 81.0; 82.3 -> 79.3 | Hurt |
Closest papers by meaning
- Simultaneous Feature Aggregating and Hashing for Large-Scale Image Search (CVPR 2017 · similarity 0.87)
- Deep Supervised Hashing for Fast Image Retrieval (CVPR 2016 · similarity 0.87)
- Learning Super-Features for Image Retrieval (ICLR 2022 · similarity 0.87)
- Towards Optimal Fine Grained Retrieval via Decorrelated Centralized Loss with Normalize-Scale Layer (AAAI 2019 · similarity 0.87)
- Exploiting Distribution Constraints for Scalable and Efficient Image Retrieval (ICLR 2025 · similarity 0.86)
- QuARI: Query Adaptive Retrieval Improvement (NeurIPS 2025 · similarity 0.85)
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.