Research map / Video / Action recognition
ViViT: A Video Vision Transformer
Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, Cordelia Schmid
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 |
|---|---|---|---|
| use tubelet embedding with central frame init instead of filter inflation (ablation) | ViViT-B Model 1 · Kinetics 400 | top-1 accuracy: 77.6 -> 79.2 | Improved |
| use tubelet embedding with central frame init instead of uniform frame sampling (ablation) | ViViT-B Model 1 · Kinetics 400 | top-1 accuracy: 78.5 -> 79.2 | Improved |
| initialise tubelet embedding from pretrained image model rather than randomly (ablation) | ViViT-B Model 1 · Kinetics 400 | top-1 accuracy: 73.2 -> 79.2 | Improved |
| add a temporal transformer (L_t=4) over frame representations instead of average pooling (ablation) | ViViT-B Model 2 · Kinetics 400 | top-1 accuracy: 75.8 -> 78.8 | Improved |
| add a temporal transformer instead of average pooling (ablation) | ViViT-B Model 2 · Epic Kitchens | action accuracy: 38.8 -> 43.7 | Improved |
| use unfactorised spatio-temporal attention instead of factorised encoder (ablation) | ViViT-B · Kinetics 400 | top-1 accuracy: 78.8 -> 80.0 | Improved |
| use factorised encoder instead of unfactorised attention on a small dataset (ablation) | ViViT-B · Epic Kitchens | action accuracy: 43.1 -> 43.7 | Improved |
| increase number of temporal transformer layers L_t from 1 to 12 (ablation) | ViViT Model 2 · Kinetics 400 | top-1 accuracy: 78.6 -> 78.9 | No clear effect |
| add progressive regularisation (stochastic depth, random augment, label smoothing, mixup, K400 init) (ablation) | ViViT Factorised encoder · Epic Kitchens | top-1 accuracy: 38.4 -> 43.7 | Improved |
| add the same regularisation set (ablation) | ViViT Factorised encoder · Something-Something v2 | top-1 accuracy: +5% | Improved |
| add label smoothing (ablation) | ViViT Factorised encoder · Epic Kitchens | top-1 accuracy: 41.1 -> 43.1 | Improved |
| increase crop size from 224 to 320 (ablation) | ViViT-L/16x2 Model 1 · Kinetics 400 | accuracy: 80.3 -> 81.0 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- AssembleNet-101 — AssembleNet: Searching for Multi-Stream Neural Connectivity in Video Architectures (ICLR 2020) · better on 1 of 1 shared result
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
- Optimizing Factorized Encoder Models: Time and Memory Reduction for Scalable and Efficient Action Recognition (ECCV 2024 · similarity 0.91)
- Video Transformers: A Survey (TPAMI 2023 · similarity 0.91)
- VidTr: Video Transformer Without Convolutions (ICCV 2021 · similarity 0.90)
- Vision Transformers Need Registers (ICLR 2024 · similarity 0.89)
- TGBFormer: Transformer-GraphFormer Blender Network for Video Object Detection (AAAI 2025 · similarity 0.89)
- Adapting Short-Term Transformers for Action Detection in Untrimmed Videos (CVPR 2024 · similarity 0.89)
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.
