Research map / Video / Video segmentation and tracking
A Large-Scale Study on Unsupervised Spatiotemporal Representation Learning
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 clips from different temporal locations rather than same-time spatial crops (ablation) | BYOL · K400 | linear accuracy: 60.6 -> 65.8 | Helped |
| Increase the number of temporal clips per video from 2 to 4 (ablation) | BYOL · K400 | linear accuracy: 65.8 -> 68.9 | Helped |
| Use a momentum encoder | Video representation learning methods · K400 | linear accuracy: ~4% | Helped |
| Add negative samples across videos | Video representation learning methods · K400 and UCF101 | accuracy | No clear effect |
| Allow positives up to 60 seconds apart (ablation) | BYOL on IG-Curated-1M · K400 | linear accuracy: 62.7 -> 64.1 | Helped |
| Increase clips per iteration instead of extending training from 50 to 200 epochs (ablation) | BYOL on IG-Curated-1M · K400 | linear accuracy: 64.1 -> 67.7; 64.1 -> 60.2 | Helped |
| Combine stronger color and cropping augmentations (ablation) | MoCo, ρ=4 · K400 | linear accuracy: 67.8 -> 69.0 | Helped |
| Use cosine annealing for momentum (ablation) | MoCo · K400 | linear accuracy: 64.5 -> 65.5 | Helped |
| Increase projection MLP depth from 2 to 3 layers (ablation) | MoCo · K400 | linear accuracy: 64.6 -> 65.8 | Helped |
| Increase projection MLP depth from 2 to 3 layers (ablation) | BYOL · UCF101 | accuracy: 92.7 -> 92.0 | Hurt |
| Increase predictor depth from 2 to 3 layers (ablation) | BYOL · K400 | linear accuracy: 65.8 -> 63.6 | Hurt |
| Use supervised rather than unsupervised K400 pre-training | BYOL · AVA | mAP: 22.2 -> 23.4 | Hurt |
Closest papers by meaning
- 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)
- Learning Transferable Spatiotemporal Representations From Natural Script Knowledge (CVPR 2023 · similarity 0.88)
- Self-Supervised Spatiotemporal Representation Learning by Exploiting Video Continuity (AAAI 2022 · similarity 0.88)
- UniSTD: Towards Unified Spatio-Temporal Learning across Diverse Disciplines (CVPR 2025 · similarity 0.87)
- Wavelet-Driven Spatiotemporal Predictive Learning: Bridging Frequency and Time Variations (AAAI 2024 · similarity 0.86)
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.