Research map / Detection and segmentation / CNN features
Rethinking ImageNet Pre-Training
Kaiming He, Ross Girshick, Piotr Dollar
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 ImageNet pre-training (ablation) | Mask R-CNN R50-FPN GN, 2x schedule · COCO val2017 | AP bbox: 36.8 -> 40.3 | Improved |
| add ImageNet pre-training (ablation) | Mask R-CNN R50-FPN GN, 6x schedule · COCO val2017 | AP bbox: 41.3 -> 41.1 | No clear effect |
| add ImageNet pre-training (ablation) | Mask R-CNN R101-FPN GN, 6x schedule · COCO val2017 | AP bbox: 42.7 -> 42.2 | No clear effect |
| train from scratch for 6x instead of 2x iterations (ablation) | Mask R-CNN R50-FPN GN · COCO val2017 | AP bbox: 36.8 -> 41.3 | Improved |
| add ImageNet pre-training at AP75 (R50) (ablation) | Mask R-CNN R50-FPN GN best schedule · COCO val2017 | AP bbox 75: 44.6 -> 45.6 | Hurt |
| add ImageNet pre-training to ResNeXt-152 (ablation) | Mask R-CNN X152 GN · COCO val2017 | AP bbox / AP mask: 50.9/43.2 -> 50.3/42.5 | No clear effect |
| add ImageNet pre-training with test-time augmentation (ablation) | Mask R-CNN R50 · COCO val2017 | AP mask: 41.6 -> 40.9 | No clear effect |
| add ImageNet pre-training for keypoint detection (ablation) | Mask R-CNN R50-FPN GN · COCO keypoints val2017 | keypoint AP: 65.6 -> 65.5 | No clear effect |
| add ImageNet pre-training with 35k COCO images (ablation) | Mask R-CNN R50-FPN GN · COCO 35k train images | AP bbox: 36.3 -> 36.3 | No clear effect |
| add ImageNet pre-training with 10k COCO images (ablation) | Mask R-CNN R50-FPN GN · COCO 10k train images | AP bbox: 25.9 -> 26.0 | No clear effect |
| add ImageNet pre-training with 3.5k COCO images (ablation) | Mask R-CNN R50-FPN GN · COCO 3.5k train images | AP bbox: 9.3 -> 16.0 | Improved |
| add ImageNet pre-training with 1k COCO images (ablation) | Mask R-CNN R50-FPN GN · COCO 1k train images | AP bbox: 3.5 -> 9.9 | Improved |
Closest papers by meaning
- Rethinking Pre-training and Self-training (NeurIPS 2020 · similarity 0.94)
- PreDet: Large-Scale Weakly Supervised Pre-Training for Detection (ICCV 2021 · similarity 0.90)
- Cheaper Pre-training Lunch: An Efficient Paradigm for Object Detection (ECCV 2020 · similarity 0.89)
- The Lottery Tickets Hypothesis for Supervised and Self-Supervised Pre-Training in Computer Vision Models (CVPR 2021 · similarity 0.89)
- Learning Rich Features at High-Speed for Single-Shot Object Detection (ICCV 2019 · similarity 0.89)
- Aligning Pretraining for Detection via Object-Level Contrastive Learning (NeurIPS 2021 · similarity 0.89)
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.
