Research map / Detection and segmentation / CNN features
Learning From Noisy Anchors for One-Stage Object Detection
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 |
|---|---|---|---|
| Use cleanliness-based soft labels and sample re-weighting | RetinaNet with ResNet-101 · COCO test-dev2017 | AP: 39.1 -> 41.8/43.4 | Helped |
| Use cleanliness-based soft labels and sample re-weighting | RetinaNet with ResNeXt-101-32×8d · COCO test-dev2017 | AP: 40.8 -> 44.1/45.5 | Helped |
| Apply cleanliness-based soft labels and re-weighting to the RPN | Faster R-CNN RPN · COCO minival | AR100: 43.3 -> 51.1 | Helped |
| Apply cleanliness-based training to the RPN | Faster R-CNN with FPN · COCO minival | AP: 36.8 -> 37.8 | Helped |
| Add cleanliness-based soft labels without sample re-weighting (ablation) | RetinaNet, ResNet-50 · COCO minival | AP: 36.2 -> 37.1 | Helped |
| Add cleanliness-based sample re-weighting while retaining hard labels (ablation) | RetinaNet, ResNet-50 · COCO minival | AP: 36.2 -> 36.7 | Helped |
| Combine cleanliness-based soft labels with sample re-weighting (ablation) | RetinaNet, ResNet-50 · COCO minival | AP: 36.2 -> 37.7 | Helped |
| Set re-weighting exponent γ to 1.75 (ablation) | Cleanliness-based RetinaNet · COCO minival | AP: 37.7 -> 35.9 | Hurt |
| Increase the positive candidate count from N=30 to N=80 (ablation) | Cleanliness-based RetinaNet · COCO minival | AP: 37.7 -> 36.6 | Hurt |
| Increase localization contribution α from 0.5 to 0.75 (ablation) | Cleanliness-based RetinaNet · COCO minival | AP: 37.7 -> 38.0 | Helped |
| Use only localization accuracy (α=1.0), rather than balancing localization and classification (ablation) | Cleanliness-based RetinaNet · COCO minival | AP: 38.0 -> 37.8 | Hurt |
Closest papers by meaning
- Bridging the Gap Between Anchor-Based and Anchor-Free Detection via Adaptive Training Sample Selection (CVPR 2020 · similarity 0.88)
- FreeAnchor: Learning to Match Anchors for Visual Object Detection (NeurIPS 2019 · similarity 0.87)
- Dynamic Anchor Learning for Arbitrary-Oriented Object Detection (AAAI 2021 · similarity 0.87)
- Multiple Anchor Learning for Visual Object Detection (CVPR 2020 · similarity 0.86)
- A Dual Weighting Label Assignment Scheme for Object Detection (CVPR 2022 · similarity 0.86)
- Object Detection Benchmarks are Incomplete: The Role of Label Errors and Annotation Uncertainty (NeurIPS 2026 · 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.