Research map / Video / Action recognition
Shifted Chunk Transformer for Spatio-Temporal Representational 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 |
|---|---|---|---|
| add one shifted MSA layer (1 shifted frame) in place of standard MSA (ablation) | SCT-S · Kinetics-400 | Top-1 accuracy (%): 76.91 -> 78.41 | Helped |
| add one shifted MSA layer (1 shifted frame) in place of standard MSA (ablation) | SCT-S · UCF101 | Top-1 accuracy (%): 97.01 -> 98.02 | Helped |
| use 5 shifted frames instead of 1 (ablation) | SCT-S · Kinetics-400 | Top-1 accuracy (%): 78.41 -> 77.02 | Hurt |
| use 2 shifted MSA layers instead of 1 (ablation) | SCT-S · Kinetics-400 | Top-1 accuracy (%): 78.41 -> 77.45 | Hurt |
| replace ViLT with ViT-B-16 frame encoder (16x16 patches) (ablation) | SCT-S · Kinetics-400 | Top-1 accuracy (%): 78.41 -> 75.33 | Hurt |
| use ViT-L-16 instead of SCT-L as frame encoder (ablation) | SCT-L · Kinetics-400 | Top-1 accuracy (%): 83.02 -> 79.15 | Hurt |
| reduce ViT-B patch size from 16 to 8 (ablation) | ViT-B-16 · Kinetics-400 | Top-1 accuracy (%): 75.3 -> 78.4 | Helped |
| reduce ViT-B patch size from 16 to 8 (ablation) | ViT-B-16 · UCF101 | Top-1 accuracy (%): 95.3 -> 97.0 | Helped |
| replace ViLT with residual convolution block (ablation) | SCT-S · Kinetics-400 | Top-1 accuracy (%): 78.4 -> 75.1 | Hurt |
| replace ViLT with image LSH attention (ablation) | SCT-S · Kinetics-400 | Top-1 accuracy (%): 78.4 -> 76.6 | Hurt |
| remove ViLT (ablation) | SCT-S · Kinetics-400 | Top-1 accuracy (%): 78.4 -> 63.2 | Hurt |
| remove LSH attention (ablation) | SCT-S · Kinetics-400 | Top-1 accuracy (%): 78.4 -> 76.2 | Hurt |
Closest papers by meaning
- UniFormer: Unified Transformer for Efficient Spatial-Temporal Representation Learning (ICLR 2022 · similarity 0.87)
- Learning Transferable Spatiotemporal Representations From Natural Script Knowledge (CVPR 2023 · similarity 0.86)
- SSTVOS: Sparse Spatiotemporal Transformers for Video Object Segmentation (CVPR 2021 · similarity 0.85)
- Compositional Video Understanding with Spatiotemporal Structure-based Transformers (CVPR 2024 · similarity 0.85)
- UniSTD: Towards Unified Spatio-Temporal Learning across Diverse Disciplines (CVPR 2025 · similarity 0.84)
- Adapting Short-Term Transformers for Action Detection in Untrimmed Videos (CVPR 2024 · similarity 0.84)
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.