Research map / Label-efficient and robust learning / Recognition and long-tail
Google Landmarks Dataset v2 - A Large-Scale Benchmark for Instance-Level Recognition and Retrieval
Tobias Weyand, Andre Araujo, Bingyi Cao, Jack Sim
Check 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 |
|---|---|---|---|
| Train ResNet101+ArcFace on GLDv2-train-clean instead of GLDv1-train | ResNet101+ArcFace · Revisited Oxford and Revisited Paris | mAP (%): 68.9 -> 76.2; 83.4 -> 87.3; 45.3 -> 55.6; 67.2 -> 74.2 | Improved |
| Train ResNet101+ArcFace on GLDv2-train-no-tail instead of GLDv2-train-clean (ablation) | ResNet101+ArcFace · Revisited Oxford and Revisited Paris | mAP (%): 76.2 -> 76.1; 87.3 -> 86.6; 55.6 -> 55.1; 74.2 -> 72.5 | Hurt |
| Use GLDv2-train-clean rather than GLDv1-train for ResNet101+ArcFace | GLDv2 recognition · Recognition testing | μAP (%): 33.25 -> 28.56 | Hurt |
| Use GLDv2-train-clean rather than GLDv1-train for ResNet101+ArcFace | GLDv2 retrieval · Retrieval testing | mAP@100 (%): 20.67 -> 25.57 | Improved |
| Add local feature spatial verification to DELG global descriptors | DELG · GLDv2 recognition testing | μAP (%): 32.03 -> 58.45 | Improved |
| Add local feature spatial verification to DELG global descriptors | DELG · GLDv2 retrieval testing | mAP@100 (%): 21.71 -> 24.54 | Improved |
| Use ArcFace rather than Triplet or AP loss | ResNet101 global descriptor · GLDv2 retrieval testing | mAP@100 (%): 18.71 (AP); 18.94 (Triplet); 21.35 (CosFace); 25.57 (ArcFace) | Improved |
| Re-annotate missing recognition labels | GLDv2 recognition challenge methods · GLDv2 recognition testing | μAP (%): 35.54 -> 69.39; 37.61 -> 66.53; 35.99 -> 53.08 | Improved |
| Re-annotate retrieval relevance labels | GLDv2 retrieval challenge methods · GLDv2 retrieval testing | Precision@100 (%): 4.73 -> 6.09; 4.83 -> 6.42; 3.97 -> 5.13 | Improved |
| Re-annotate retrieval relevance labels | GLDv2 retrieval challenge methods · GLDv2 retrieval testing | mAP@100 (%): 37.25 -> 37.19; 34.75 -> 34.38; 32.18 -> 32.10 | No clear effect |
| Use GLDv2-train-clean rather than Landmarks-full or Landmarks-clean | ResNet101+ArcFace · Revisited Oxford and Revisited Paris | mAP (%): Landmarks-full: 50.8, 70.4, 24.3, 47.1; Landmarks-clean: 54.2, 70.7, 28.3, 46.0; GLDv2-train-clean: 76.2, 87.3, 55.6, 74 | Improved |
Closest papers by meaning
- Detect-To-Retrieve: Efficient Regional Aggregation for Image Search (CVPR 2019 · similarity 0.89)
- On Train-Test Class Overlap and Detection for Image Retrieval (CVPR 2024 · similarity 0.88)
- Does Progress On Object Recognition Benchmarks Improve Generalization on Crowdsourced, Global Data? (ICLR 2024 · similarity 0.87)
- GeoDE: a Geographically Diverse Evaluation Dataset for Object Recognition (NeurIPS 2023 · similarity 0.86)
- From Keypoints to Object Landmarks via Self-Training Correspondence: A novel approach to Unsupervised Landmark Discovery (TPAMI 2023 · similarity 0.86)
- Instance-Level Composed Image Retrieval (NeurIPS 2025 · 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.
