Research map / Label-efficient and robust learning / Self-supervised and contrastive
Scaling and Benchmarking Self-Supervised Visual Representation Learning
Priya Goyal, Dhruv Mahajan, Abhinav Gupta, Ishan Misra
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 |
|---|---|---|---|
| scale pre-training data from 1M to 100M images (ablation) | Jigsaw and Colorization (AlexNet, ResNet-50) · VOC07 classification | mAP | Improved |
| use ResNet-50 instead of AlexNet as pre-training data grows (ablation) | Jigsaw and Colorization · VOC07 classification | mAP | Improved |
| increase Jigsaw permutations from 100 to 10k (ResNet-50) (ablation) | Jigsaw, YFCC-1M · VOC07 classification | mAP: 5 point mAP improvement | Improved |
| increase Jigsaw permutations from 100 to 10k (AlexNet) (ablation) | Jigsaw, YFCC-1M · VOC07 classification | mAP: 1.9 point improvement | Improved |
| vary Colorization soft-encoding neighbours K (ablation) | Colorization, YFCC-1M · VOC07 classification | mAP: ~2 point mAP variation | No clear effect |
| pre-train Jigsaw on ImageNet-22k instead of ImageNet-1k (ablation) | Jigsaw ResNet-50 · VOC07 detection (frozen conv body) | mAP: 56.6 -> 67.1 | Improved |
| pre-train Jigsaw on ImageNet-22k instead of YFCC-100M (ablation) | Jigsaw ResNet-50 · VOC07 detection (frozen conv body) | mAP: 62.3 -> 67.1 | Improved |
| pre-train Jigsaw on YFCC-100M instead of ImageNet-22k (ablation) | Jigsaw ResNet-50 · Places205 linear (layer4) | top-1 accuracy: 41.5 -> 44.8 | Improved |
| pre-train Jigsaw on ImageNet-22k instead of YFCC-100M (ablation) | Jigsaw ResNet-50 · VOC07 linear SVM (layer4) | mAP: 71.0 -> 71.9 | No clear effect |
| replace random-init (scratch) with Jigsaw YFCC-100M pre-training | ResNet-50 PSPNet · NYUv2 surface normals | mean angle error: 26.3 -> 22.4 | Improved |
| replace Places205 supervised with Jigsaw YFCC-100M | ResNet-50 PSPNet · NYUv2 surface normals | pixels within 11.25 degrees: 41.8 -> 44.6 | Improved |
| replace ImageNet-1k supervised with Jigsaw ImageNet-22k | ResNet-50 · VOC07 linear SVM (layer4) | mAP: 80.4 -> 71.9 | Hurt |
Closest papers by meaning
- Revisiting Self-Supervised Visual Representation Learning (CVPR 2019 · similarity 0.92)
- How Well Do Self-Supervised Models Transfer? (CVPR 2021 · similarity 0.92)
- Towards Efficient and Effective Self-Supervised Learning of Visual Representations (ECCV 2022 · similarity 0.91)
- When Does Contrastive Visual Representation Learning Work? (CVPR 2022 · similarity 0.90)
- Probing the Mid-level Vision Capabilities of Self-Supervised Learning (CVPR 2025 · similarity 0.89)
- Masked Scene Modeling: Narrowing the Gap Between Supervised and Self-Supervised Learning in 3D Scene Understanding (CVPR 2025 · 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.
