Research map / Video / Video segmentation and tracking
UniFormer: Unified Transformer for Efficient Spatial-Temporal Representation Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Unify local and global relation aggregation in LLGG (ablation) | UniFormer · Kinetics-400 1×4 | Top-1 accuracy: 77.2 -> 79.3 | Helped |
| Add shallow-layer transformer-style FFN blocks instead of MobileNet blocks (ablation) | UniFormer · Kinetics-400 1×4 | Top-1 accuracy: 78.6 -> 79.3 | Helped |
| Use joint rather than divided spatiotemporal attention (ablation) | UniFormer deep layers · Kinetics-400 1×4 | Top-1 accuracy: 78.7 -> 79.3 | Helped |
| Add dynamic position embedding (ablation) | UniFormer · ImageNet-1K | Top-1 accuracy: +0.5% | Helped |
| Add dynamic position embedding (ablation) | UniFormer · Kinetics-400 | Top-1 accuracy: +1.7% | Helped |
| Add global MHRA to local-only stages, progressing from LLLL toward LLGG (ablation) | UniFormer · Kinetics-400 1×4 | Top-1 accuracy: 77.2 -> 79.3 | Helped |
| Replace local MHRA throughout with global MHRA (ablation) | UniFormer · Kinetics-400 1×4 | Top-1 accuracy: 79.3 -> 75.3 | Hurt |
| Increase local tube size from 5 to 7 (ablation) | UniFormer · Kinetics-400 1×4 | Top-1 accuracy: 79.3 -> 79.1 | No clear effect |
| Increase local tube size from 5 to 9 (ablation) | UniFormer · Kinetics-400 1×4 | Top-1 accuracy: 79.3 -> 78.9 | Hurt |
| Pretrain LLGG on Kinetics-400 rather than ImageNet (ablation) | UniFormer LLGG · Something-Something V1 | Top-1 accuracy: +1.8 | Helped |
| Pretrain LLLL on Kinetics-400 rather than ImageNet (ablation) | UniFormer LLLL · Something-Something V1 | Top-1 accuracy: +0.0 | No clear effect |
| Pretrain divided-attention LLGG on Kinetics-400 rather than ImageNet (ablation) | UniFormer with divided attention · Something-Something V1 | Top-1 accuracy: -0.1 | Hurt |
Closest papers by meaning
- UniSTD: Towards Unified Spatio-Temporal Learning across Diverse Disciplines (CVPR 2025 · similarity 0.87)
- Spatiotemporal Fusion in 3D CNNs: A Probabilistic View (CVPR 2020 · similarity 0.87)
- Shifted Chunk Transformer for Spatio-Temporal Representational Learning (NeurIPS 2021 · similarity 0.87)
- Compositional Video Understanding with Spatiotemporal Structure-based Transformers (CVPR 2024 · similarity 0.86)
- LaVin-DiT: Large Vision Diffusion Transformer (CVPR 2025 · similarity 0.85)
- TGBFormer: Transformer-GraphFormer Blender Network for Video Object Detection (AAAI 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.