Research map / Video / Video segmentation and tracking
Contrastive Spatio-Temporal Pretext Learning for Self-Supervised Video Representation
Read the paperCheck 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 |
|---|---|---|---|
| Add STOR to BYOL with playback-rate and rotation prediction (ablation) | CSTP on R(2+1)D · UCF-101 split 1 | accuracy: 84.4 -> 85.6 | Helped |
| Add STOR to BYOL with playback-rate and rotation prediction (ablation) | CSTP on R(2+1)D · HMDB-51 split 1 | accuracy: 54.1 -> 55.1 | Helped |
| Add spatial and temporal overlap prediction together (ablation) | R(2+1)D trained from scratch with base augmentation · UCF-101 split 1 | accuracy: 60.3 -> 69.6 | Helped |
| Add spatial overlap prediction (ablation) | R(2+1)D trained from scratch with base augmentation · UCF-101 split 1 | accuracy: 60.3 -> 67.8 | Helped |
| Add temporal overlap prediction (ablation) | R(2+1)D trained from scratch with base augmentation · UCF-101 split 1 | accuracy: 60.3 -> 66.5 | Helped |
| Use 5 rather than 2 overlap-rate candidates (ablation) | STOR pre-training with R(2+1)D · UCF-101 split 1 | accuracy: 72.5 -> 76.2 | Helped |
| Use 5 rather than 2 overlap-rate candidates (ablation) | STOR pre-training with R(2+1)D · HMDB-51 split 1 | accuracy: 42.5 -> 45.4 | Helped |
| Increase STOR overlap-rate candidates from 5 to 6 (ablation) | STOR pre-training with R(2+1)D · UCF-101 split 1 | accuracy: 76.2 -> 75.4 | Hurt |
| Add playback-rate augmentation to base augmentation (ablation) | R(2+1)D trained from scratch · UCF-101 split 1 | accuracy: 60.3 -> 67.5 | Helped |
| Add playback-rate prediction to playback-rate augmentation (ablation) | R(2+1)D trained from scratch · UCF-101 split 1 | accuracy: 67.5 -> 77.2 | Helped |
| Add playback-rate, rotation, and STOR pretext tasks to BYOL (ablation) | BYOL-based CSTP on R(2+1)D · UCF-101 split 1 | accuracy: 76.6 -> 85.6 | Helped |
| Use BYOL rather than SimSiam as the contrastive framework (ablation) | CSTP with all three pretext tasks · UCF-101 split 1 | accuracy: 83.9 -> 85.6 | Helped |
Closest papers by meaning
- Contextualized Spatio-Temporal Contrastive Learning With Self-Supervision (CVPR 2022 · similarity 0.90)
- Spatiotemporal Contrastive Video Representation Learning (CVPR 2021 · similarity 0.89)
- Self-Supervised Spatio-Temporal Representation Learning for Videos by Predicting Motion and Appearance Statistics (CVPR 2019 · similarity 0.89)
- No More Shortcuts: Realizing the Potential of Temporal Self-Supervision (AAAI 2024 · similarity 0.84)
- State Space Prompting via Gathering and Spreading Spatio-Temporal Information for Video Understanding (NeurIPS 2025 · similarity 0.84)
- Video Language Model Pretraining with Spatio-temporal Masking (CVPR 2025 · similarity 0.84)
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.