Research map / Label-efficient and robust learning / Self-supervised and contrastive
Uncovering the Hidden Dynamics of Video Self-supervised Learning under Distribution Shifts
Pritam Sarkar, Ahmad Beirami, Ali Etemad
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 |
|---|---|---|---|
| finetune instead of linear evaluation | VSSL video encoders · InD average | accuracy: 63.2% -> 71.9% | Improved |
| finetune instead of linear evaluation | VSSL video encoders · OoD average | accuracy: 31.5% -> 37.4% | Improved |
| contrastive pretraining (negatives) vs supervised on multiple concurrent shifts | V-SimCLR vs V-Supervised · Sims4Action top-down+synthetic, finetuned | accuracy: 44.3 -> 63.1 | Improved |
| use masked autoencoding instead of contrastive | V-MAE vs V-SimCLR · CharadesEgo ego viewpoint, linear | mAP: 12.7 -> 10.9 | Hurt |
| use V-MAE on low-resolution surveillance | V-MAE vs V-Supervised · TinyVirat-v2 FT OoD | top-1 accuracy: 24.5 -> 29.1 | Improved |
| finetune HMDB-trained model vs frozen encoder (reversed source shift) | V-BYOL · HMDB to UCF source shift | accuracy | Hurt |
| end-to-end finetune contrastive encoders under temporal perturbations | contrastive/non-contrastive VSSL · temporally perturbed inputs | robustness | Hurt |
| use frozen contrastive encoders for open-set | V-SimCLR vs V-DINO · Open-set UCF101/HMDB51 linear | AUC: 79.4 -> 49.9 | Hurt |
| finetune contrastive models for open-set recognition | V-MAE vs V-SimCLR · Open-set K400/UCF finetuned | AUC: 55.1 -> 63.0 | Improved |
| aggressive cropping [0.08,1.0] for contrastive methods (ablation) | contrastive VSSL · UCF101 linear SVM | accuracy | Improved |
| aggressive cropping for masked autoencoding (ablation) | V-MAE · UCF101 | accuracy | Hurt |
| remove temporal information (single frames) (ablation) | V-MAE · UCF101 linear | accuracy: 76.2 -> 5.6 | Hurt |
Closest papers by meaning
- No More Shortcuts: Realizing the Potential of Temporal Self-Supervision (AAAI 2024 · similarity 0.87)
- How Severe Is Benchmark-Sensitivity in Video Self-Supervised Learning? (ECCV 2022 · similarity 0.87)
- Self-Supervised Learning of Video-Induced Visual Invariances (CVPR 2020 · similarity 0.87)
- Self-Supervised Video Representation Learning in a Heuristic Decoupled Perspective (IJCV 2026 · similarity 0.87)
- TransRank: Self-Supervised Video Representation Learning via Ranking-Based Transformation Recognition (CVPR 2022 · similarity 0.87)
- Self-Supervised Learning of Compressed Video Representations (ICLR 2021 · 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.
