Research map / Video / Action recognition
Selective Feature Compression for Efficient Activity Recognition Inference
Chunhui Liu, Xinyu Li, Hao Chen, Davide Modolo, Joseph Tighe
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 |
|---|---|---|---|
| replace 30-crop dense sampling with SFC | Slow-Only I3D-50 8x8 · Kinetics-400 | top-1 accuracy: 74.4 -> 74.6 | No clear effect |
| replace 30-crop dense sampling with SFC | R(2+1)D-152 · Kinetics-400 | top-1 accuracy: 79.4 -> 80.0 | Improved |
| replace 30-crop dense sampling with SFC | Slow-Only I3D-50 8x8 · Kinetics-400 | FLOPS: 1643G -> 247G | Improved |
| use SFC with tau=2 instead of no compression (tau=1) (ablation) | SFC I3D-50 8x8 · Kinetics-400 (I3D-50 8x8) | top-1 accuracy: 73.8 -> 74.6 | Improved |
| compress more aggressively (tau=4 instead of tau=2) (ablation) | SFC I3D-50 8x8 · Kinetics-400 (I3D-50 8x8) | top-1 accuracy: 74.6 -> 72.0 | Hurt |
| insert SFC after Res3 instead of after Res1 (ablation) | SFC I3D-50 8x8 · Kinetics-400 (I3D-50 8x8) | top-1 accuracy: 73.1 -> 74.6 | Improved |
| use TopK pooling instead of average pooling (ablation) | SFC I3D-50 8x8 · Kinetics-400 (I3D-50 8x8) | top-1 accuracy: 72.9 -> 74.6 | Improved |
| use TopK pooling instead of max pooling (ablation) | SFC I3D-50 8x8 · Kinetics-400 (I3D-50 8x8) | top-1 accuracy: 72.6 -> 74.6 | Improved |
| add abstraction layer on q and k only (value untouched) (ablation) | SFC I3D-50 8x8 · Kinetics-400 (I3D-50 8x8) | top-1 accuracy: 72.7 -> 74.6 | Improved |
| also transform value v with the abstraction layer (ablation) | SFC I3D-50 8x8 · Kinetics-400 (I3D-50 8x8) | top-1 accuracy: 74.6 -> 63.2 | Hurt |
| use 3x1x1 temporal conv instead of 3x3x3 (ablation) | SFC I3D-50 8x8 · Kinetics-400 (I3D-50 8x8) | top-1 accuracy: 72.0 -> 74.6 | Improved |
| use 3x1x1 conv instead of 1x1x1 (ablation) | SFC I3D-50 8x8 · Kinetics-400 (I3D-50 8x8) | top-1 accuracy: 72.5 -> 74.6 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- SCSampler — SCSampler: Sampling Salient Clips From Video for Efficient Action Recognition (ICCV 2019) · better on 0 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Reconstruction-Free Action Inference from Compressive Imagers (TPAMI 2016 · similarity 0.87)
- Action Recognition With Spatial-Temporal Discriminative Filter Banks (ICCV 2019 · similarity 0.86)
- MGSampler: An Explainable Sampling Strategy for Video Action Recognition (ICCV 2021 · similarity 0.86)
- Compressed Video Action Recognition (CVPR 2018 · similarity 0.86)
- Action Detail Matters: Refining Video Recognition with Local Action Queries (CVPR 2025 · similarity 0.86)
- Temporal Action Detection Model Compression by Progressive Block Drop (CVPR 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.
