Research map / Video / Video segmentation and tracking
UniFormerV2: Unlocking the Potential of Image ViTs for Video Understanding
Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Limin Wang, Yu Qiao
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 global UniBlock to CLIP-ViT baseline (ablation) | UniFormerV2 · Kinetics-400 | top-1 accuracy: 83.1 -> 84.4 | Improved |
| add local UniBlock to CLIP-ViT baseline (ablation) | UniFormerV2 · Something-Something V2 | top-1 accuracy: 45.1 -> 67.7 | Improved |
| add temporal downsampling with doubled frames (ablation) | UniFormerV2 · Something-Something V2 | top-1 accuracy: 68.7 -> 69.5 | Improved |
| use local MHRA instead of ST-Adapter (ablation) | local UniBlock · Something-Something V2 | top-1 accuracy: 68.0 -> 69.1 | Improved |
| use convolution instead of temporal self-attention for local temporal modeling (ablation) | local UniBlock · Something-Something V2 | top-1 accuracy: 65.2 -> 67.5 | Improved |
| add DPE to local MHRA (ablation) | local UniBlock · Something-Something V2 | top-1 accuracy: 69.1 -> 69.1 | No clear effect |
| use sequential fusion rather than parallel or hierarchical KV (ablation) | multi-stage fusion · Something-Something V2 | top-1 accuracy: 69.1/68.9 -> 69.5 | Improved |
| use more queries (4 or 16) in global block (ablation) | global UniBlock · Something-Something V2 | top-1 accuracy: 69.5 -> 69.1 (4 queries), 68.6 (16 queries) | Hurt |
| post-pretrain on Kinetics-710 before finetuning (ablation) | UniFormerV2 · Kinetics-400 | top-1 accuracy: 84.4 -> 85.6 | Improved |
| K710 training instead of K400+K600+K700 co-training (ablation) | UniFormerV2 · Kinetics-700 | top-1 accuracy: -0.2 for co-training vs +0.5 for K710 | Improved |
| apply UniFormerV2 design to different pretrained ViTs (ablation) | image ViTs · Kinetics-400 / SSV2 | top-1 accuracy: CLIP 84.4 / 69.5 vs TimeSformer 78.7 / 59.5 | Improved |
| UniFormerV2-L versus VideoMAE-L on SSV2 | UniFormerV2-L · Something-Something V2 | top-1 accuracy: 73.0 vs 74.3 | Hurt |
Closest papers by meaning
- UniFormer: Unified Transformer for Efficient Spatial-Temporal Representation Learning (ICLR 2022 · similarity 0.89)
- ViViT: A Video Vision Transformer (ICCV 2021 · similarity 0.89)
- Adapting Short-Term Transformers for Action Detection in Untrimmed Videos (CVPR 2024 · similarity 0.87)
- Memory Consolidation Enables Long-Context Video Understanding (ICML 2024 · similarity 0.87)
- TTVidT: Decoupling the Temporal Axis for Efficient Motion-Centric Video Pretraining (NeurIPS 2026 · similarity 0.86)
- Vision Transformers Need Registers (ICLR 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.
