Research map / Video / Video segmentation and tracking
AttentionNAS: Spatiotemporal Attention Cell Search for Video Classification
Xiaofang Wang, Xuehan Xiong, Maxim Neumann, AJ Piergiovanni, Michael S. Ryoo, Anelia Angelova, Kris M. Kitani, Wei Hua
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 |
|---|---|---|---|
| insert searched attention cells instead of non-local blocks | I3D · Kinetics-600 | top-1 accuracy: 76.87 -> 77.86 | Improved |
| insert searched attention cells instead of non-local blocks | S3D · Kinetics-600 | top-1 accuracy: 77.56 -> 78.51 | Improved |
| insert searched cells instead of non-local blocks on MiT | I3D · MiT | top-1 accuracy: 28.54 -> 29.58 | Improved |
| use position-specific instead of position-agnostic cell designs | I3D · Kinetics-600 | top-1 accuracy: 77.56 -> 77.86 | Improved |
| use position-specific instead of position-agnostic cell designs (ablation) | I3D · MiT | top-1 accuracy: 28.18 -> 29.58 | Improved |
| use position-specific instead of position-agnostic cell designs (ablation) | S3D · MiT | top-1 accuracy: 29.19 -> 29.82 | Improved |
| search with differentiable method instead of GPB, transfer to optical flow | S3D · Kinetics-600 (optical flow) | top-1 accuracy: 67.02 -> 66.29 | Hurt |
| transfer cells found on RGB to optical flow | I3D · Kinetics-600 (optical flow) | top-1 accuracy: 61.14 -> 66.81 | Improved |
| insert cells searched for S3D into I3D | I3D · Kinetics-600 | top-1 accuracy: 75.58 -> 77.81 | Improved |
| insert I3D cells into I3D-R50 | I3D-R50 · MiT | top-1 accuracy: 30.63 -> 32.48 | Improved |
| transfer cells searched on Kinetics-600 to MiT with GPB | I3D · MiT | top-1 accuracy: 27.38 -> 27.62 | No clear effect |
| add map-based attention to the supergraph alongside dot-product (ablation) | I3D · Kinetics-600 | top-1 accuracy: 77.86 -> 77.82 | No clear effect |
Closest papers by meaning
- Recurrent Mixture Density Network for Spatiotemporal Visual Attention (ICLR 2017 · similarity 0.89)
- Rethinking Spatiotemporal Feature Learning: Speed-Accuracy Trade-offs in Video Classification (ECCV 2018 · similarity 0.88)
- Interaction-aware Spatio-temporal Pyramid Attention Networks for Action Classification (ECCV 2018 · similarity 0.88)
- Attention Clusters: Purely Attention Based Local Feature Integration for Video Classification (CVPR 2018 · similarity 0.87)
- Class Semantics-Based Attention for Action Detection (ICCV 2021 · similarity 0.86)
- Faster Neighborhood Attention: Reducing the O(n^2) Cost of Self Attention at the Threadblock Level (NeurIPS 2024 · 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.
