Research map / Detection and segmentation / CNN features
MetaAnchor: Learning to Detect Objects with Customized Anchors
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 |
|---|---|---|---|
| Replace predefined anchor functions with functions generated from customized priors | RetinaNet · COCO-mini, 3×3 anchors, threshold 0.5/0.4 | mmAP: 26.5 -> 26.9 | Helped |
| Replace predefined anchor functions with functions generated from customized priors | RetinaNet · COCO minival | mmAP: 35.8 -> 37.5 | Helped |
| Apply MetaAnchor to the YOLOv2 detector | YOLOv2 · COCO minival | mmAP: 18.9 -> 21.2 | Helped |
| Add greedy search to choose inference anchors (ablation) | MetaAnchor · COCO-mini | mmAP: 27.0 -> 27.7 | Helped |
| Increase inference anchor density from 3×3 to 9×9 (ablation) | MetaAnchor · COCO-mini | mmAP: 0.7∼1.1% | Helped |
| Use data-dependent rather than data-independent anchor generation (ablation) | MetaAnchor · COCO-mini, 5×5 anchors | mmAP: 27.3 -> 27.3 | No clear effect |
| Use data-dependent rather than data-independent anchor generation with searched inference anchors (ablation) | MetaAnchor · COCO-mini | mmAP: 27.6 -> 28.0 | Helped |
| Train with stricter positive/negative IoU thresholds (0.6/0.5) instead of 0.5/0.4 (ablation) | RetinaNet baseline · COCO-mini | mmAP: 26.5 -> 25.7 | Hurt |
| Train with stricter IoU thresholds while using MetaAnchor (ablation) | MetaAnchor · COCO-mini, 5×5 anchors | mmAP: 26.9 -> 27.3 | Helped |
| Omit a range of ground-truth box sizes during training (ablation) | MetaAnchor · COCO-mini, 5×5 anchors; evaluated on all validation data | mmAP: 27.1 -> 23.0; RetinaNet: 26.9 -> 20.8 | Mixed |
| Use searched inference anchors instead of fixed 3×3 anchors (ablation) | MetaAnchor · VOC 2007 test (without sofa), trained on COCO-full | mAP: 83.1 -> 83.3 | Helped |
| Add MetaAnchor to RetinaNet | RetinaNet · PASCAL VOC 2007 | mAP@0.5: 80.3->80.6 | Helped |
Closest papers by meaning
- FreeAnchor: Learning to Match Anchors for Visual Object Detection (NeurIPS 2019 · similarity 0.90)
- Multiple Anchor Learning for Visual Object Detection (CVPR 2020 · similarity 0.89)
- Bridging the Gap Between Anchor-Based and Anchor-Free Detection via Adaptive Training Sample Selection (CVPR 2020 · similarity 0.88)
- Region Proposal by Guided Anchoring (CVPR 2019 · similarity 0.87)
- Improving Single Domain-Generalized Object Detection: A Focus on Diversification and Alignment (CVPR 2024 · similarity 0.84)
- DetKDS: Knowledge Distillation Search for Object Detectors (ICML 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.