Research map / Video / Video segmentation and tracking
TokenLearner: Adaptive Space-Time Tokenization for Videos
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 |
|---|---|---|---|
| insert TokenLearner (16 tokens) after layer 12 of ViViT-L/16 (ablation) | ViViT-L/16 · Kinetics-400 | top-1 accuracy / GFLOPS: 83.4 -> 83.5; 1446 -> 766 GFLOPS | Helped |
| insert TokenLearner (8 tokens) later, at layer 18 (ablation) | ViViT-L/16 · Kinetics-400 | top-1 accuracy: 83.4 -> 84.5 | Helped |
| use TokenLearner with L/10 backbone and 16 tokens | ViViT · Kinetics-400 | top-1 accuracy: 84.5 -> 85.4 | Helped |
| add TokenFuser to TokenLearner (L/10) (ablation) | ViViT-L/10 TokenLearner · Kinetics-600 | top-1 accuracy: 86.1 -> 86.3 | Helped |
| replace full joint space-time MHSA with TokenLearner (ablation) | X(2+1)D-M bottleneck · Charades 12 fps | mAP: 64.0 -> 63.4; with Conv2D 62.5 -> 66.3 | Helped |
| use TokenLearner instead of joint space-time MHSA on AViD (ablation) | X(2+1)D-M bottleneck · AViD | accuracy: 53.3 -> 53.8 | Helped |
| replace TokenLearner with disjoint spatial+temporal transformers (ablation) | X(2+1)D-M bottleneck · Charades 6 fps | mAP: 59.6 -> 58.0 | Hurt |
| replace learned spatial-attention tokens with a fixed grid of 8 tokens (ablation) | TokenLearner · Charades | accuracy: 59.6 -> 58.8 | Hurt |
| generate tokens directly with a dense layer instead of spatial attention (ablation) | TokenLearner · Charades | accuracy: 59.6 -> 56.6 | Hurt |
| generate tokens by global pooling and MLP instead of spatial attention (ablation) | TokenLearner · Charades | accuracy: 59.6 -> 58.6 | Hurt |
| add Conv2D before TokenLearner (ablation) | TokenLearner · Charades 12 fps | mAP: 63.4 -> 66.3 | Helped |
| add Conv2D before TokenLearner (ablation) | TokenLearner · AViD | accuracy: 53.8 -> 53.7 | No clear effect |
Closest papers by meaning
- Single-pass Adaptive Image Tokenization for Minimum Program Search (NeurIPS 2025 · similarity 0.90)
- ElasticTok: Adaptive Tokenization for Image and Video (ICLR 2025 · similarity 0.89)
- Video Token Merging for Long Video Understanding (NeurIPS 2024 · similarity 0.89)
- VQToken: Neural Discrete Token Representation Learning for Extreme Token Reduction in Video Large Language Models (NeurIPS 2025 · similarity 0.88)
- Adaptive Length Image Tokenization via Recurrent Allocation (ICLR 2025 · similarity 0.88)
- Efficient Long Video Tokenization via Coordinate-based Patch Reconstruction (CVPR 2025 · 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.