Research map / Video / Action recognition
Frame-Wise Action Representations for Long Videos via Sequence Contrastive 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 sequence contrastive loss instead of corresponding-frame-only contrastive learning (ablation) | CARL on PennAction · PennAction | Classification: 88.05 -> 93.07 | Helped |
| Use sequence contrastive loss instead of corresponding-frame-only contrastive learning (ablation) | CARL on PennAction · PennAction | Kendall's Tau: 0.891 -> 0.985 | Helped |
| Add a Transformer encoder to the ResNet-50 frame encoder (ablation) | CARL on PennAction · PennAction | Classification: 68.63 -> 93.07 | Helped |
| Add a Transformer encoder rather than two 3D convolutional layers (ablation) | CARL on PennAction · PennAction | Progress: 0.705 -> 0.918 | Helped |
| Use 3 Transformer layers rather than 1 (ablation) | CARL on PennAction · PennAction | Classification: 92.15 -> 93.07 | Helped |
| Use 4 Transformer layers rather than 3 (ablation) | CARL on PennAction · PennAction | Progress: 0.918 -> 0.910 | Hurt |
| Train ResNet-50 Block 5 rather than freezing all layers (ablation) | CARL on PennAction · PennAction | Classification: 90.63 -> 93.07 | Helped |
| Train Blocks 4 and 5 instead of only Block 5 (ablation) | CARL on PennAction · PennAction | Classification: 93.07 -> 92.98 | Hurt |
| Use Gaussian variance σ²=1 rather than σ²=10 (ablation) | CARL on PennAction · PennAction | Classification: 93.07 -> 92.95 | Hurt |
| Use Gaussian variance σ²=25 rather than σ²=10 (ablation) | CARL on PennAction · PennAction | Classification: 93.07 -> 92.03 | Hurt |
| Use temperature τ=1.0 rather than τ=0.1 (ablation) | CARL on PennAction · PennAction | Classification: 93.07 -> 91.57 | Hurt |
| Use random sampling rather than even sampling (ablation) | CARL · FineGym99 | Fine-grained action classification: 38.44 -> 41.75 | Helped |
Closest papers by meaning
- Self-Supervised Video Representation Learning via Latent Time Navigation (AAAI 2023 · similarity 0.90)
- Learning by Aligning Videos in Time (CVPR 2021 · similarity 0.90)
- Self-Supervised Video Representation Learning by Context and Motion Decoupling (CVPR 2021 · similarity 0.89)
- Cycle-Contrast for Self-Supervised Video Representation Learning (NeurIPS 2020 · similarity 0.89)
- Cross-Architecture Self-Supervised Video Representation Learning (CVPR 2022 · similarity 0.89)
- Storyboard-guided Alignment for Fine-grained Video Action Recognition (NeurIPS 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.