Research map / Detection and segmentation / CNN features
Cheaper Pre-training Lunch: An Efficient Paradigm for Object Detection
Dongzhan Zhou, Xinchi Zhou, Hongwen Zhang, Shuai Yi, Wanli Ouyang
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 |
|---|---|---|---|
| replace ImageNet pre-training with Montage pre-training at 1/4 cost | Faster R-CNN C4 ResNet-50 · COCO val2017 | AP: 34.8 -> 36.3 | Improved |
| replace ImageNet pre-training with Montage pre-training | Faster R-CNN FPN ResNet-50 · COCO val2017 | AP: 36.2 -> 36.5 | No clear effect |
| replace ImageNet pre-training with Montage pre-training | Mask R-CNN FPN ResNet-50 · COCO val2017 | AP: 37.3 -> 37.5 | No clear effect |
| replace ImageNet pre-training with Montage pre-training | Faster R-CNN ResNet-101 · COCO val2017 | AP: 38.3 -> 39.2 | Improved |
| replace ImageNet pre-training with Montage pre-training | Faster R-CNN ResNeXt101-32x4d · COCO val2017 | AP: 40.2 -> 40.2 | No clear effect |
| use Montage pre-training instead of training detector from scratch (GN), 1x schedule | Faster R-CNN ResNet-50 · COCO val2017 | AP: 28.6 -> 36.3 | Improved |
| use Montage pre-training instead of training detector from scratch (GN), 2x schedule | Faster R-CNN ResNet-50 · COCO val2017 | AP: 32.6 -> 37.5 | Improved |
| add Montage assembly and ERF-adaptive loss to MS-COCO classification pre-training (1.73 GPU days) (ablation) | MS-COCO pre-training · COCO val2017 | AP: 33.5 -> 36.3 | Improved |
| extend MS-COCO classification pre-training to ImageNet cost without Montage (ablation) | MS-COCO w/o Montage · COCO val2017 | AP: 33.5 -> 34.4 | No clear effect |
| use pad and crop rather than warp to fit samples (ablation) | Montage pre-training · COCO val2017 | AP: 34.6 -> 35.2 | Improved |
| use pad and crop rather than resize to fit samples (ablation) | Montage pre-training · COCO val2017 | AP: 34.7 -> 35.2 | Improved |
| use ERF-adaptive dense classification instead of global classification (ablation) | Montage pre-training · COCO val2017 | AP: 34.3 -> 36.3 | Improved |
Closest papers by meaning
- Rethinking ImageNet Pre-Training (ICCV 2019 · similarity 0.89)
- PreDet: Large-Scale Weakly Supervised Pre-Training for Detection (ICCV 2021 · similarity 0.88)
- AlignDet: Aligning Pre-training and Fine-tuning in Object Detection (ICCV 2023 · similarity 0.87)
- Towards All-in-One Pre-Training via Maximizing Multi-Modal Mutual Information (CVPR 2023 · similarity 0.87)
- Rethinking Pre-training and Self-training (NeurIPS 2020 · similarity 0.87)
- Learning Rich Features at High-Speed for Single-Shot Object Detection (ICCV 2019 · similarity 0.87)
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.
