Research map / Detection and segmentation / CNN features
Revisiting Unreasonable Effectiveness of Data in Deep Learning Era
Chen Sun, Abhinav Shrivastava, Saurabh Singh, Abhinav Gupta
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 |
|---|---|---|---|
| pretrain on JFT-300M then fine-tune instead of random init | ResNet-101 · ImageNet val | top-1 accuracy: 77.5 -> 79.2 | Improved |
| pretrain on JFT-300M instead of ImageNet | ResNet-101 Faster RCNN · COCO test-dev detection | mAP@[0.5,0.95]: 34.3 -> 36.7 | Improved |
| add ImageNet initialization to JFT-300M pretraining | ResNet-101 Faster RCNN · COCO test-dev detection | mAP@[0.5,0.95]: 36.7 -> 37.4 | Improved |
| pretrain on JFT-300M instead of ImageNet | ResNet-101 Faster RCNN · PASCAL VOC 2007 test detection | mAP@0.5: 76.3 -> 81.4 | Improved |
| add ImageNet initialization to JFT-300M pretraining | ResNet-101 Faster RCNN · PASCAL VOC 2007 test detection | mAP@0.5: 81.4 -> 81.3 | No clear effect |
| pretrain on JFT-300M instead of ImageNet | DeepLab-ASPP-L · PASCAL VOC 2012 val segmentation | mIoU: 73.6 -> 75.3 | Improved |
| add ImageNet initialization to JFT-300M pretraining | DeepLab-ASPP-L · PASCAL VOC 2012 val segmentation | mIoU: 75.3 -> 76.5 | Improved |
| pretrain on JFT-300M instead of ImageNet | G-RMI pose ResNet-101 · COCO test-dev pose estimation | AP: 62.4 -> 64.8 | Improved |
| add ImageNet initialization to JFT-300M pretraining | G-RMI pose ResNet-101 · COCO test-dev pose estimation | AP: 64.8 -> 64.4 | No clear effect |
| pretrain JFT-300M for more epochs (1.3 to 4) (ablation) | ResNet-101 Faster RCNN · COCO minival | mAP@[0.5,0.95]: 35.0 -> 36.8 | Improved |
| pretrain ImageNet for more epochs (3 to 150) (ablation) | ResNet-101 Faster RCNN · COCO minival | mAP@[0.5,0.95]: 22.2 -> 34.5 | Improved |
| use 18K JFT classes instead of 1K ImageNet-matching classes (ablation) | 30M-image JFT pretraining · COCO minival | mAP@[0.5,0.95]: 31.2 -> 31.9 | No clear effect |
Closest papers by meaning
- Big Transfer (BiT): General Visual Representation Learning (ECCV 2020 · similarity 0.88)
- When Do We Not Need Larger Vision Models? (ECCV 2024 · similarity 0.87)
- Rethinking the Inception Architecture for Computer Vision (CVPR 2016 · similarity 0.86)
- Large Scale Visual Recognition through Adaptation using Joint Representation and Multiple Instance Learning (JMLR 2016 · similarity 0.86)
- Sequential Modeling Enables Scalable Learning for Large Vision Models (CVPR 2024 · similarity 0.85)
- Scalable Pre-training of Large Autoregressive Image Models (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.
