Research map / Detection and segmentation / CNN features
On Train-Test Class Overlap and Detection for Image Retrieval
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Removes landmark classes from Google Landmarks v2 clean that overlap with Revisited Oxford/Paris (yielding RGLDv2-clean) and proposes CiDeR, a single-stage, end-to-end detect-to-retrieve model that needs no location supervision.
Problem
Training sets for instance-level image retrieval (GLDv2-clean) overlap in landmark classes with the standard evaluation sets ROxford/RParis, which inflates reported performance and makes comparisons unfair; also, detect-to-retrieve methods typically need two-stage training and location supervision.
Approach
Identify and remove landmark categories in GLDv2-clean that overlap with ROxford/RParis via retrieval, spatial verification and human inspection to create RGLDv2-clean; and an attentional localization module (thresholded, learned-weight spatial attention masks with randomized background value) trained end-to-end with the global descriptor, requiring no location supervision.
Key results
- Training on RGLDv2-clean instead of GLDv2-clean causes mAP drops of roughly 1% to 30% across reproduced SOTA methods, and changes method rankings (e.g., Token drops 18.2 mAP on the average difference column).
- CiDeR-FT reaches 79.4 mean mAP on SfM-120k-trained evaluation in Table 5 and 75.3 mean mAP on RGLDv2-clean in Table 6, outperforming DOLG (67.9) and GLAM (65.6) on the same training set.
- Without fine-tuning, CiDeR improves over DOLG on RGLDv2-clean by 1.3% mAP on ROxf+R1M (medium) and 5.1% on ROxf+R1M (hard).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train on RGLDv2-clean instead of GLDv2-clean | reproduced retrieval methods · ROxf and RPar | mAP: 1% up to 30% | Hurt |
| Use CiDeR instead of the strongest listed baseline | ROxf and RPar medium and hard protocols, trained on RGLDv2-clean · ROxf and RPar medium and hard | mAP: 73.7 vs 72.2; 79.4 vs 84.7; 54.9 vs 49.6; 68.5 vs 65.6 | Helped |
| Fine-tune the backbone in CiDeR | CiDeR · ROxf and RPar medium and hard | mAP: 77.8 vs 73.7; 79.4 vs 79.4; 61.9 vs 54.9; 75.3 vs 68.5 | Helped |
| Add selective context (ablation) | ResNet101 with GeM pooling · Ox5k | mAP: 80.2 -> 87.6 | Helped |
| Add attentional localization to selective context (ablation) | ResNet101 with GeM pooling and selective context · Ox5k | mAP: 87.6 -> 89.4 | Helped |
| Add backbone enhancement to selective context (ablation) | ResNet101 with GeM pooling and selective context · Ox5k | mAP: 87.6 -> 88.2 | Helped |
| Add backbone enhancement to selective context and attentional localization (ablation) | CiDeR without backbone enhancement · Ox5k | mAP: 89.4 -> 89.9 | Helped |
| Use randomized background β ~ N(0.1, 0.5) rather than fixed β=0.0 (ablation) | attentional localization · Ox5k | mAP: 87.4 -> 90.2 | Helped |
| Use randomized background β ~ N(0.1, 0.9) rather than fixed β=0.0 (ablation) | attentional localization · Ox5k | mAP: 87.4 -> 89.9 | Helped |
| Use 2 masks rather than 1 (ablation) | attentional localization · Ox5k | mAP: 87.5 -> 89.9 | Helped |
| Use 2 masks rather than 3 (ablation) | attentional localization · Ox5k | mAP: 89.4 vs 89.9 | No clear effect |
| Use 2 masks rather than 6 (ablation) | attentional localization · Ox5k | mAP: 89.4 vs 89.9 | No clear effect |
Surprising findings
- Only 1,565 images (about 0.1%) from 18 categories are removed, yet performance drops sharply and rankings of methods change.
- Top-5 retrievals from GLDv2-clean for ROxford/RParis queries all depict the same landmark, whereas NC-clean and SfM-120k retrievals show none.
Limitations
- Overlap removal relies on automatic matching plus human inspection; the removal process is limited to ROxford and RParis evaluation landmarks.
- Reported performance depends on reproduced baselines, with some methods using the authors' own code and results from the original papers for others.
Experimental setup
Datasets: GLDv2-clean RGLDv2-clean NC-clean SfM-120k Revisited Oxford Revisited Paris Oxford5k Paris6k R1M distractors
Benchmarks: Oxford5k Paris6k Revisited Oxford Revisited Paris ROxf+R1M RParis+R1M
Metrics: mAP mP@10
Baselines: GLAM DOLG SOLAR Token Yokoo et al. DIR+RPN AGeM R-ASMK
Builds on: GeM pooling ArcFace GLAM DOLG SOLAR DIR
Closest papers by meaning
- Instance-Level Composed Image Retrieval (NeurIPS 2025 · similarity 0.85)
- Let All Be Whitened: Multi-Teacher Distillation for Efficient Visual Retrieval (AAAI 2024 · similarity 0.84)
- ILIAS: Instance-Level Image retrieval At Scale (CVPR 2025 · similarity 0.84)
- LOCORE: Image Re-ranking with Long-Context Sequence Modeling (CVPR 2025 · similarity 0.84)
- Search and Detect: Training-Free Long Tail Object Detection via Web-Image Retrieval (CVPR 2025 · similarity 0.84)
- Exploiting Distribution Constraints for Scalable and Efficient Image Retrieval (ICLR 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.