Research map / Video / Video segmentation and tracking
Deformable Video Transformer
Jue Wang, Lorenzo Torresani
Check 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 |
|---|---|---|---|
| Replace fixed MHPA with D-ST+MS-A | MViT-B · Kinetics-400 | Top-1 accuracy: 78.4% -> 79.0% | Improved |
| Replace trajectory attention with D-ST+MS-A | ViT-B · Kinetics-400, 16×224² clips | Top-1 accuracy: 79.7% -> 80.5% | Improved |
| Replace MHPA with D-ST+MS-A | MViT-B · Something-Something-V2 | Top-1 accuracy: 64.7% -> 67.5% | Improved |
| Replace trajectory attention with D-ST+MS-A | ViT-B · EPIC-KITCHENS-100 Action, 16×224² clips | accuracy: 43.1% -> 43.8% | Improved |
| Replace divided space-time attention with D-ST+MS-A | ViT-B · Diving-48, same input resolution | Top-1 accuracy: 78.0% -> 86.6% | Improved |
| Use motion displacements rather than averaged pooled RGB, RGB residuals, or RGB differences (ablation) | DVT D-ST-A · Kinetics-400 and Something-Something-V2 | accuracy: K400: 72.3% -> 77.9%; 73.2% -> 77.9%; 73.6% -> 77.9%. SSv2: 62.9% -> 65.6%; 63.3% -> 65.6%; 63.5% -> 65.6%. | Improved |
| Use optical flow instead of motion displacements (ablation) | DVT D-ST-A · Kinetics-400 and Something-Something-V2 | accuracy: K400: 77.9% -> 78.2%; SSv2: 65.6% -> 66.0%. | Improved |
| Add multi-scale deformable attention to space-time deformable attention (ablation) | DVT MViT-B · Kinetics-400 | accuracy: 77.9% -> 78.1% (linear); 77.9% -> 79.0% (MLP-Mixer) | Improved |
| Use MLP-Mixer instead of linear projection for attention fusion (ablation) | DVT D-ST+MS-A, MViT-B · Kinetics-400 | accuracy: 78.1% -> 79.0% | Improved |
| Use MLP-Mixer instead of linear projection for attention fusion (ablation) | DVT D-ST+MS-A, MViT-B · Something-Something-V2 | accuracy: 66.8% -> 67.5% | Improved |
| Increase attended patches from N=4 to N=8 (ablation) | DVT D-ST+MS-A · Kinetics-400 and Something-Something-V2 | accuracy | Improved |
| Increase the number of sub-clips to B=8 (ablation) | DVT D-ST-A · Something-Something-V2 | accuracy | Hurt |
Closest papers by meaning
- Keeping Your Eye on the Ball: Trajectory Attention in Video Transformers (NeurIPS 2021 · similarity 0.90)
- Video Transformers: A Survey (TPAMI 2023 · similarity 0.89)
- ViViT: A Video Vision Transformer (ICCV 2021 · similarity 0.88)
- Video Motion Transfer with Diffusion Transformers (CVPR 2025 · similarity 0.87)
- Adapting Short-Term Transformers for Action Detection in Untrimmed Videos (CVPR 2024 · similarity 0.87)
- Understanding Video Transformers via Universal Concept Discovery (CVPR 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.
