Research map / Video / Video segmentation and tracking
DualFormer: Local-Global Stratified Transformer for Efficient Video Recognition
Yuxuan Liang, Pan Zhou, Roger Zimmermann, Shuicheng Yan
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 GP-MSA to LW-MSA in all stages (ablation) | DualFormer-T · Kinetics-400 | top-1 accuracy: 78.4 -> 79.5 | Improved |
| add LW-MSA to GP-MSA in all stages (ablation) | DualFormer-T · Kinetics-400 | top-1 accuracy: 77.6 -> 79.5 | Improved |
| add GP-MSA to early stages rather than late stages (ablation) | DualFormer-T · Kinetics-400 | top-1 accuracy: 78.8 -> 79.3 | Improved |
| use two pyramid scales instead of only (4,4,4) (ablation) | DualFormer-T GP-MSA · Kinetics-400 | top-1 accuracy: 78.4 -> 79.5 | Improved |
| use two pyramid scales instead of only (8,7,7) (ablation) | DualFormer-T GP-MSA · Kinetics-400 | top-1 accuracy: 79.3 -> 79.5 | Improved |
| enlarge LW-MSA window from 4x7x7 to 8x7x7 (32x2 input) (ablation) | DualFormer-T · Kinetics-400 | top-1 accuracy: 79.1 -> 79.5 (FLOPs 224 -> 240) | Improved |
| enlarge LW-MSA window from 8x7x7 to 16x7x7 (32x2 input) (ablation) | DualFormer-T · Kinetics-400 | top-1 accuracy: 79.5 -> 79.7 (FLOPs 240 -> 272) | Improved |
| use 32 input frames instead of 16 (ablation) | DualFormer-T · Kinetics-400 | top-1 accuracy: 78.4 -> 79.5 | Improved |
| use DWConv instead of AvgPool for pyramid downsampling (ablation) | DualFormer-T · Kinetics-400 | top-1 accuracy: 78.7 -> 79.5 | Improved |
| use DWConv instead of standard Conv for downsampling (ablation) | DualFormer-T · Kinetics-400 | top-1 accuracy: 79.5 -> 79.5 (Param 27.6 -> 21.8) | No clear effect |
| add DWConv-based PEG (ablation) | DualFormer-T · Kinetics-400 | top-1 accuracy: 78.9 -> 79.5 | Improved |
| use PEG instead of absolute position encoding (ablation) | DualFormer-T · Kinetics-400 | top-1 accuracy: 79.2 -> 79.5 | Improved |
Closest papers by meaning
- Space-time Mixing Attention for Video Transformer (NeurIPS 2021 · similarity 0.90)
- Focal Attention for Long-Range Interactions in Vision Transformers (NeurIPS 2021 · similarity 0.88)
- TGBFormer: Transformer-GraphFormer Blender Network for Video Object Detection (AAAI 2025 · similarity 0.87)
- Video Swin Transformer (CVPR 2022 · similarity 0.87)
- Pale Transformer: A General Vision Transformer Backbone with Pale-Shaped Attention (AAAI 2022 · similarity 0.87)
- An Efficient Spatio-Temporal Pyramid Transformer for Action Detection (ECCV 2022 · similarity 0.87)
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.
