Research map / Detection and segmentation / Open-set and few-shot detection
AlignDet: Aligning Pre-training and Fine-tuning in Object Detection
Ming Li, Jie Wu, Xionghui Wang, Chen Chen, Jie Qin, Xuefeng Xiao, Rui Wang, Min Zheng, Xin Pan
Read the paperCheck 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 |
|---|---|---|---|
| Add AlignDet box-domain pre-training to Mask R-CNN backbones | Mask R-CNN ResNet-50 · COCO 1x | AP^bb: 38.3 -> 39.4 | Improved |
| Add AlignDet to SOTA CBNet v2 | CBNet v2 Swin-L · COCO | AP^bb: 57.3 -> 57.8 | Improved |
| Add AlignDet pre-training and transfer to VOC | DETR · Pascal VOC | AP: 52.1 -> 58.2 | Improved |
| Add AlignDet pre-training in low-data regime with long schedule | Mask R-CNN · COCO 5% data, 90k iterations | mAP: +1.4 mAP | Improved |
| Compare AlignDet 1x against baseline trained 4x longer (ablation) | Mask R-CNN ResNet-50 · COCO | AP^bb: 39.2 -> 39.4 | Improved |
| Pre-train box-domain on COCO versus random init (ablation) | Faster R-CNN ResNet-50 · COCO 1x | AP: 37.9 -> 39.0 | Improved |
| Use COCO instead of ImageNet Subset for box-domain pre-training (ablation) | Faster R-CNN ResNet-50 · COCO 1x | AP: 38.7 -> 39.0 | Improved |
| Pre-train only the RPN while keeping Head random (ablation) | Faster R-CNN ResNet-50 · COCO 1x | AP: 38.4 -> 38.2 | Hurt |
| Pre-train the Neck (ablation) | Faster R-CNN ResNet-50 · COCO 1x | AP: 37.9 -> 38.4 | Improved |
| Add both classification and regression pretext tasks (ablation) | Faster R-CNN ResNet-50 · COCO 1x | AP: 37.9 -> 39.0 | Improved |
| Freeze backbone during box-domain pre-training (ablation) | Faster R-CNN ResNet-50 · COCO | AP: 35.5 -> 37.3 | Improved |
| Use AlignDet with 12 epochs on COCO versus SoCo on ImageNet | Mask R-CNN ResNet-50 · COCO 1x | AP: AlignDet+SwAV 42.3 vs SoCo 43.0 | No clear effect |
Closest papers by meaning
- PreDet: Large-Scale Weakly Supervised Pre-Training for Detection (ICCV 2021 · similarity 0.91)
- Adapting Object Detectors via Selective Cross-Domain Alignment (CVPR 2019 · similarity 0.89)
- DSOD: Learning Deeply Supervised Object Detectors From Scratch (ICCV 2017 · similarity 0.89)
- TOOD: Task-Aligned One-Stage Object Detection (ICCV 2021 · similarity 0.89)
- Improving Single Domain-Generalized Object Detection: A Focus on Diversification and Alignment (CVPR 2024 · similarity 0.88)
- MatchDet: A Collaborative Framework for Image Matching and Object Detection (AAAI 2024 · similarity 0.88)
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.
