Research map / Video / Video segmentation and tracking
BEVT: BERT Pretraining of Video Transformers
Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Yu-Gang Jiang, Luowei Zhou, Lu Yuan
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 video-stream masked modeling to image-stream pretraining | BEVT-I · SSv2 | Top-1 accuracy: 69.2 -> 70.6 | Improved |
| Add video-stream masked modeling to image-stream pretraining | BEVT-I · Diving-48 | Top-1 accuracy: 81.2 -> 86.7 | Improved |
| Add video-stream masked modeling to image-stream pretraining | BEVT-I · K400 | Top-1 accuracy: 80.5 -> 80.6 | No clear effect |
| Initialize the video stream with image-stream pretrained weights (ablation) | BEVT-V · SSv2 | Top-1 accuracy: 67.1 -> 70.0 | Improved |
| Initialize the video stream with image-stream pretrained weights (ablation) | BEVT-V · Diving-48 | Top-1 accuracy: 83.7 -> 85.2 | Improved |
| Initialize the video stream with image-stream pretrained weights (ablation) | BEVT-V · K400 | Top-1 accuracy: 76.2 -> 79.6 | Improved |
| Initialize joint BEVT with image-stream pretrained weights (ablation) | BEVT · SSv2 | Top-1 accuracy: 67.9 -> 70.6 | Improved |
| Initialize joint BEVT with image-stream pretrained weights (ablation) | BEVT · Diving-48 | Top-1 accuracy: 85.1 -> 86.7 | Improved |
| Initialize joint BEVT with image-stream pretrained weights (ablation) | BEVT · K400 | Top-1 accuracy: 78.5 -> 80.6 | Improved |
| Use ImageNet images rather than K400 frames for the image stream (ablation) | BEVT joint pretraining · SSv2 | Top-1 accuracy: 70.3 -> 70.6 | Improved |
| Use ImageNet images rather than K400 frames for the image stream (ablation) | BEVT joint pretraining · Diving-48 | Top-1 accuracy: 86.6 -> 86.7 | Improved |
| Use ImageNet images rather than K400 frames for the image stream (ablation) | BEVT joint pretraining · K400 | Top-1 accuracy: 80.5 -> 80.6 | Improved |
Later papers that compared with this method
2 later accepted papers report BEVT as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- AdaMAE: Adaptive Masking for Efficient Spatiotemporal Learning With Masked Autoencoders (CVPR 2023 · better on 0 of 1 shared result)
- VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training (NeurIPS 2022 · 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
- Video Transformers: A Survey (TPAMI 2023 · similarity 0.88)
- Long-Short Temporal Contrastive Learning of Video Transformers (CVPR 2022 · similarity 0.88)
- Masked Video Distillation: Rethinking Masked Feature Modeling for Self-Supervised Video Representation Learning (CVPR 2023 · similarity 0.87)
- ViViT: A Video Vision Transformer (ICCV 2021 · similarity 0.87)
- Adapting Short-Term Transformers for Action Detection in Untrimmed Videos (CVPR 2024 · similarity 0.86)
- Video In-context Learning: Autoregressive Transformers are Zero-Shot Video Imitators (ICLR 2025 · 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.
