Research map / Detection and segmentation / Open-set and few-shot detection
How Well Do Self-Supervised Models Transfer?
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use the strongest self-supervised representation instead of supervised ImageNet pretraining | Many-shot recognition transfer · Kornblith datasets | accuracy | Helped |
| Use supervised pretraining rather than self-supervised pretraining | Few-shot transfer · Kornblith datasets similar to ImageNet | accuracy | Helped |
| Use self-supervised pretraining rather than supervised pretraining | Few-shot transfer · CD-FSL datasets | accuracy | Helped |
| Select the representation by ImageNet accuracy | Transfer to recognition on ImageNet-like datasets · Many-shot and low-domain-shift few-shot recognition | transfer performance correlation | Helped |
| Select the representation by ImageNet accuracy | Transfer to unstructured or spatially sensitive tasks · ISIC, detection, surface normals, and semantic segmentation | transfer performance correlation | Hurt |
| Use SimCLR-v2 features instead of supervised features | Frozen-backbone object detection · Pascal VOC | AP: 51.99 -> 54.95 | Helped |
| Use BYOL features instead of supervised features | Finetuned object detection · Pascal VOC | AP: 53.26 -> 54.91 | Helped |
| Use SimCLR-v2 features instead of supervised features | Surface normal estimation · NYUv2 | mean angular error: 33.52 -> 28.77 | Helped |
| Use PCL-v1 features instead of supervised features | Semantic segmentation · ADE20K | mean IoU: 0.2563 -> 0.2983 | Helped |
| Use supervised features rather than self-supervised features | Feature inversion and image reconstruction · Recognition datasets | perceptual distance: 0.47 vs 0.49 | Helped |
| Use self-supervised features rather than supervised features | Unscaled linear transfer · Many-shot recognition | expected calibration error: 10.35 vs 7.05 | Helped |
| Use self-supervised features rather than supervised features | Fine-tuned transfer · Many-shot recognition | attentive diffusion | Mixed |
Closest papers by meaning
- Revisiting Self-Supervised Visual Representation Learning (CVPR 2019 · similarity 0.90)
- When Does Contrastive Visual Representation Learning Work? (CVPR 2022 · similarity 0.89)
- Boosting Self-Supervised Learning via Knowledge Transfer (CVPR 2018 · similarity 0.89)
- Self-Supervised Learning of Video-Induced Visual Invariances (CVPR 2020 · similarity 0.89)
- Is ImageNet worth 1 video? Learning strong image encoders from 1 long unlabelled video (ICLR 2024 · similarity 0.88)
- Probing the Mid-level Vision Capabilities of Self-Supervised Learning (CVPR 2025 · 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.