Research map / Label-efficient and robust learning / Self-supervised and contrastive
PASS: An ImageNet replacement for self-supervised pretraining without humans
Yuki M Asano, Christian Rupprecht, Andrew Zisserman, Andrea Vedaldi
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 PASS instead of ImageNet-1k (transfer) | MoCo-v2 · COCO bbox detection | AP: 38.7 -> 38.0 | No clear effect |
| pretrain on PASS instead of ImageNet-1k | MoCo-v2 · Linear probing ImageNet-1k | top-1 accuracy: 67.5 -> 59.5 | Hurt |
| pretrain on PASS instead of ImageNet-1k | MoCo-v2 · Linear probing Herbarium19 | top-1 accuracy: 42.3 -> 46.5 | Improved |
| pretrain on PASS instead of ImageNet-1k | MoCo-v2 · Linear probing CIFAR-10 | top-1 accuracy: 91.6 -> 81.0 | Hurt |
| remove images with humans from ImageNet pretraining (ablation) | MoCo-v2 · COCO dense pose | AP_GPS: 65.0 -> 64.8 | No clear effect |
| include humans in PASS (skip human filtering) (ablation) | MoCo-v2 on PASS · Linear probing ImageNet-1k | top-1 accuracy: 59.5 -> 59.7 | No clear effect |
| match IN-1k class statistics (~IN800 split) (ablation) | MoCo-v2 on PASS · Linear probing ImageNet-1k | top-1 accuracy: 59.5 -> 62.4 | Improved |
| match IN-1k class statistics (~IN800 split) (ablation) | MoCo-v2 on PASS · Linear probing Places205 | top-1 accuracy: 52.8 -> 51.5 | Hurt |
| use all 6M cleaned images (Random6M) instead of 1.28M balanced PASS (ablation) | MoCo-v2 · Linear probing ImageNet-1k | top-1 accuracy: 59.5 -> 61.3 | Improved |
| use DINO instead of MoCo-v2 on PASS (ablation) | PASS pretraining · 4-shot Pascal VOC07 | mAP: 57.6 -> 64.3 | Improved |
| use SwAV instead of MoCo-v2 on PASS (ablation) | PASS pretraining · Linear probing Herbarium19 | top-1 accuracy: 46.5 -> 54.4 | Improved |
| lower RRC minimum crop size from 0.2 to 0.1 (ablation) | MoCo-v2 on PASS · Linear probing Herbarium19 | top-1 accuracy: 46.5 -> 48.9 | Improved |
Closest papers by meaning
- ImageNet-21K Pretraining for the Masses (NeurIPS 2021 · similarity 0.87)
- Is ImageNet worth 1 video? Learning strong image encoders from 1 long unlabelled video (ICLR 2024 · similarity 0.87)
- Revisiting the Transferability of Supervised Pretraining: An MLP Perspective (CVPR 2022 · similarity 0.86)
- The Effectiveness of MAE Pre-Pretraining for Billion-Scale Pretraining (ICCV 2023 · similarity 0.86)
- Beyond Random Augmentations: Pretraining with Hard Views (ICLR 2025 · similarity 0.86)
- Cross-view and Cross-pose Completion for 3D Human Understanding (CVPR 2024 · similarity 0.85)
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.
