Research map / Video / Action recognition
Anticipative Video Transformer
Rohit Girdhar, Kristen Grauman
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 |
|---|---|---|---|
| add L_cls and L_feat anticipative losses (ablation) | AVT with AVT-b backbone · EK55 val (RGB, anticipative training ablation) | action top-5 accuracy: 25.9 -> 30.1 | Improved |
| add L_cls intermediate action loss (ablation) | AVT with AVT-b backbone · EK55 val (RGB, anticipative training ablation) | action top-5 accuracy: 25.9 -> 28.8 | Improved |
| add L_feat alone without L_cls (ablation) | AVT with AVT-b backbone · EK55 val (RGB, anticipative training ablation) | action top-5 accuracy: 25.9 -> 23.9 | Hurt |
| add L_feat to L_cls (ablation) | AVT with AVT-b backbone · EK55 val (RGB, anticipative training ablation) | action top-5 accuracy: 28.8 -> 30.1 | Improved |
| add L_cls and L_feat anticipative losses (ablation) | AVT-h on IG65M irCSN152 features · EK55 val (RGB, anticipative training ablation) | action top-5 accuracy: 31.4 -> 31.7 | No clear effect |
| train with anticipative losses, longer temporal context (ablation) | AVT with AVT-b backbone · EK100 val (RGB) | action recall@5: 11.2 -> 14.9 | Improved |
| train with naive loss, longer temporal context (ablation) | AVT with AVT-b backbone · EK100 val (RGB) | action recall@5: 11.0 -> 13.1 | Improved |
| replace TSN features with end-to-end AVT-b backbone (IN21+1k) | AVT-h · EK100 val (RGB) | action recall@5: 13.6 -> 14.4 | Improved |
| replace TSN features with irCSN152 IG65M features | AVT-h · EK100 val (RGB) | action recall@5: 13.6 -> 12.8 | Hurt |
| use IN21k instead of IN1k pretraining | AVT-b backbone · EK100 val (RGB) | action recall@5: 13.4 -> 14.9 | Improved |
| replace RULSTM head with AVT-h on same TSN features | AVT-h · EK100 val (RGB) | action recall@5: 13.3 -> 13.6 | Improved |
| add object modality and late fuse with RGB | AVT · EK100 val | action recall@5: 14.9 -> 15.9 | Improved |
Closest papers by meaning
- Rethinking Learning Approaches for Long-Term Action Anticipation (ECCV 2022 · similarity 0.89)
- Future Transformer for Long-Term Action Anticipation (CVPR 2022 · similarity 0.89)
- Adapting Short-Term Transformers for Action Detection in Untrimmed Videos (CVPR 2024 · similarity 0.88)
- Video Transformers: A Survey (TPAMI 2023 · similarity 0.88)
- VEDIT: Latent Prediction Architecture For Procedural Video Representation Learning (ICLR 2025 · similarity 0.86)
- ActFusion: a Unified Diffusion Model for Action Segmentation and Anticipation (NeurIPS 2024 · similarity 0.85)
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.
