Research map / Detection and segmentation / CNN features
Rethinking Classification and Localization for 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 |
|---|---|---|---|
| Assign classification to fc-head and regression to conv-head rather than using the FPN fc-head for both tasks | Double-Head, ResNet-50 · COCO val2017 | AP: 36.8 -> 39.8 | Helped |
| Assign classification to fc-head and regression to conv-head rather than using the FPN fc-head for both tasks | Double-Head, ResNet-101 · COCO val2017 | AP: 39.1 -> 41.5 | Helped |
| Use Double-Head-Ext instead of FPN | Object detector, ResNet-50 · COCO val2017 | AP: +3.5 AP | Helped |
| Use Double-Head-Ext instead of FPN | Object detector, ResNet-101 · COCO val2017 | AP: +2.8 AP | Helped |
| Use Double-Head-Ext instead of FPN | Object detector, ResNet-50 · VOC07 test | AP: +1.8 AP | Helped |
| Use Double-Head-Ext instead of FPN | Object detector, ResNet-101 · COCO test-dev | AP: 42.3 AP | Helped |
| Use fc-head for classification and conv-head for regression instead of two same-structure heads (ablation) | Double-Head, ResNet-50 · COCO val2017 | AP: 37.3 -> 38.8; 33.8 -> 38.8 | Helped |
| Assign fc-head to regression and conv-head to classification (ablation) | Double-Head-Reverse, ResNet-50 · COCO val2017 | AP: 38.8 -> 32.6 | Hurt |
| Increase conv-head depth from one residual block to two (ablation) | Double-Head, ResNet-50 · COCO val2017 | AP: 36.7 (-0.1) -> 38.7 (+1.9) | Helped |
| Add non-local blocks to the conv-head at K=5 (ablation) | Double-Head, ResNet-50 · COCO val2017 | AP: 39.5 (+2.7) -> 39.8 (+3.0) | Helped |
| Increase conv-head depth from K=5 to K=7 with non-local blocks (ablation) | Double-Head, ResNet-50 · COCO val2017 | AP: 39.8 (+3.0) -> 40.0 (+3.2) | Helped |
| Increase training from 180k to 360k iterations (ablation) | Double-Head, ResNet-50 · COCO val2017 | AP: 39.8 to 40.4; gains 0.6 AP | Helped |
Closest papers by meaning
- Revisiting the Sibling Head in Object Detector (CVPR 2020 · similarity 0.88)
- Dynamic Head: Unifying Object Detection Heads With Attentions (CVPR 2021 · similarity 0.87)
- R-FCN: Object Detection via Region-based Fully Convolutional Networks (NeurIPS 2016 · similarity 0.87)
- Correlation Loss: Enforcing Correlation between Classification and Localization (AAAI 2023 · similarity 0.86)
- Recognize Any Regions (NeurIPS 2024 · similarity 0.85)
- CAM Back Again: Large Kernel CNNs from a Weakly Supervised Object Localization Perspective (CVPR 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.