Research map / Video / Video segmentation and tracking
Enhancing Self-Supervised Video Representation Learning via Multi-Level Feature Optimization
Rui Qian, Yuxi Li, Huabin Liu, John See, Shuangrui Ding, Xian Liu, Dian Li, Weiyao Lin
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 |
|---|---|---|---|
| add combined instance+semantic graph targets for lower-level features (multi-level) (ablation) | R3D-18 baseline · UCF-101 linear probe | top-1 accuracy: 58.1 -> 63.2 | Improved |
| add combined graph targets for lower-level features (multi-level) (ablation) | R3D-18 baseline · HMDB-51 linear probe | top-1 accuracy: 28.8 -> 33.4 | Improved |
| use one-hot labels for lower-level features (ablation) | R3D-18 baseline · UCF-101 linear probe | top-1 accuracy: 58.1 -> 57.8 | Hurt |
| use instance-wise distribution only for lower-level features (ablation) | R3D-18 baseline · UCF-101 linear probe | top-1 accuracy: 58.1 -> 59.4 | Improved |
| use semantic-wise distribution only for lower-level features (ablation) | R3D-18 baseline · UCF-101 linear probe | top-1 accuracy: 58.1 -> 54.5 | Hurt |
| apply combined graph constraint to lower-level features rather than only high-level (ablation) | R3D-18 · UCF-101 linear probe | top-1 accuracy: 60.4 -> 63.2 | Improved |
| add own temporal modeling (NTM) vs none (ablation) | R3D-18 · UCF-101 linear probe | top-1 accuracy: 55.9 -> 63.2 | Improved |
| add own temporal modeling (NTM) vs none (ablation) | S3D · UCF-101 linear probe | top-1 accuracy: 53.8 -> 61.1 | Improved |
| add own temporal modeling (NTM) vs none (ablation) | SlowFast · UCF-101 linear probe | top-1 accuracy: 57.8 -> 63.9 | Improved |
| use conventional input-level temporal augmentation instead of none (ablation) | R3D-18 · UCF-101 linear probe | top-1 accuracy: 55.9 -> 43.2 | Hurt |
| use conventional input-level temporal augmentation instead of none (ablation) | S3D · UCF-101 linear probe | top-1 accuracy: 53.8 -> 41.1 | Hurt |
| use conventional input-level temporal augmentation instead of none (ablation) | SlowFast · UCF-101 linear probe | top-1 accuracy: 57.8 -> 65.4 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 2 on most shared results.
- TaCo — TACo: Token-Aware Cascade Contrastive Learning for Video-Text Alignment (ICCV 2021) · better on 2 of 2 shared results
- SpeedNet — SpeedNet: Learning the Speediness in Videos (CVPR 2020) · better on 0 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Contrast and Order Representations for Video Self-Supervised Learning (ICCV 2021 · similarity 0.89)
- Hierarchical Self-Supervised Representation Learning for Movie Understanding (CVPR 2022 · similarity 0.89)
- Self-Supervised Spatio-Temporal Representation Learning for Videos by Predicting Motion and Appearance Statistics (CVPR 2019 · similarity 0.88)
- Self-Supervised Video Representation Learning by Context and Motion Decoupling (CVPR 2021 · similarity 0.88)
- No More Shortcuts: Realizing the Potential of Temporal Self-Supervision (AAAI 2024 · similarity 0.88)
- Static and Dynamic Concepts for Self-Supervised Video Representation Learning (ECCV 2022 · 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.
