Research map / Video / Video-language understanding
Principles of Visual Tokens for Efficient Video Understanding
Xinyue Hao, Gen Li, Shreyank N Gowda, Robert B. Fisher, Jonathan Huang, Anurag Arnab, Laura Sevilla-Lara
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 |
|---|---|---|---|
| compute oracle gradient with the true label instead of the predicted label (ablation) | token oracle, 50% tokens · SS-V2 | Top-1 accuracy: 70.00 -> 78.52 | Improved |
| add adaptive budget (LITE++) to LITE90 (ablation) | VideoMAE-LITE · SS-V2 | GFLOPs / Top-1: 159x2x3 -> 105x2x3 GFLOPs; 70.54 -> 69.94 | Mixed |
| add adaptive budget (LITE++) to LITE70 (ablation) | VideoMAE-LITE · SS-V2 | GFLOPs / Top-1: 118x2x3 -> 84x2x3 GFLOPs; 70.44 -> 69.86 | Mixed |
| add adaptive budget (LITE++) to LITE30 (ablation) | VideoMAE-LITE · SS-V2 | GFLOPs / Top-1: 46x2x3 -> 40x2x3 GFLOPs; 68.33 -> 67.74 | Mixed |
| replace VideoMAE backbone with weaker TimeSformer for LITE50 | LITE · SS-V2 | Top-1 accuracy: 56.23 -> 48.28 (vs full TimeSformer-B) | Hurt |
| select tokens with LITE90 instead of using all tokens | VideoMAE · SS-V2 | Top-1 accuracy: 70.50 -> 70.54 | No clear effect |
| select tokens with LITE50 instead of using all tokens | VideoMAE · Kinetics-400 | Top-1 accuracy: 81.25 -> 80.36 | Hurt |
| use LITE instead of ToMe at similar GFLOPs | VideoMAE · SS-V2 | Top-1 accuracy: 69.63 -> 69.91 | Improved |
| use LITE instead of ToMe at similar GFLOPs | VideoMAE · Kinetics-400 | Top-1 accuracy: 79.95 -> 80.36 | Improved |
| transfer selector trained on K400 to SS-V2 (LITE90) | VideoMAE-LITE · SS-V2 | Top-1 accuracy: 70.50 -> 70.28 | No clear effect |
| transfer selector trained on K400 to AVA detection (LITE70) | VideoMAE-LITE · AVA | mAP: 31.50 -> 30.83 | No clear effect |
| use a more complex selector (transformer blocks, more layers, global information, improved oracle) (ablation) | LITE selector · SS-V2 | Top-1 accuracy | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- VideoMAE — VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training (NeurIPS 2022) · better on 1 of 3 shared results
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
- Video Token Merging for Long Video Understanding (NeurIPS 2024 · similarity 0.90)
- TokenLearner: Adaptive Space-Time Tokenization for Videos (NeurIPS 2021 · similarity 0.90)
- FlexSelect: Flexible Token Selection for Efficient Long Video Understanding (NeurIPS 2025 · similarity 0.88)
- Don't Look Twice: Faster Video Transformers with Run-Length Tokenization (NeurIPS 2024 · similarity 0.87)
- ApET: Approximation-Error Guided Token Compression for Efficient VLMs (CVPR 2026 · similarity 0.87)
- MovieChat: From Dense Token to Sparse Memory for Long Video Understanding (CVPR 2024 · 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.
