Research map / Video / Action recognition
SlowFast Networks for Video Recognition
Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, Kaiming He
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 lightweight Fast pathway with T-conv lateral connections (ablation) | Slow-only R50 · Kinetics-400 (R50 4x16) | top-1 accuracy: 72.6 -> 75.6 | Improved |
| add lateral connections (T-conv) between pathways (ablation) | SlowFast · Kinetics-400 (R50 4x16) | top-1 accuracy: 73.5 -> 75.6 | Improved |
| shrink Fast pathway channel ratio to beta=1/32 (ablation) | SlowFast · Kinetics-400 (R50 4x16) | top-1 accuracy: 72.6 -> 74.2 | Improved |
| increase beta from 1/8 to 1/4 (ablation) | SlowFast · Kinetics-400 (R50 4x16) | top-1 accuracy: 75.6 -> 75.6 | No clear effect |
| use gray-scale input for Fast pathway instead of RGB (ablation) | SlowFast · Kinetics-400 (R50 4x16) | top-1 accuracy: 75.6 -> 75.5 | No clear effect |
| use optical flow as Fast pathway input instead of RGB (ablation) | SlowFast · Kinetics-400 (R50 4x16) | top-1 accuracy: 75.6 -> 73.8 | Hurt |
| halve Fast pathway spatial resolution (beta=1/4) (ablation) | SlowFast · Kinetics-400 (R50 4x16) | top-1 accuracy: 75.6 -> 74.7 | Hurt |
| use time-difference frames as Fast pathway input (ablation) | SlowFast · Kinetics-400 (R50 4x16) | top-1 accuracy: 75.6 -> 74.5 | Hurt |
| ImageNet pre-training versus training from scratch (ablation) | 3D R-50 · Kinetics-400 | top-1 accuracy: 73.4 -> 73.5 | No clear effect |
| add Fast pathway (ablation) | Slow-only R101 16x8 vs SlowFast · Charades | mAP: 39.0 -> 42.1 | Improved |
| add non-local blocks (ablation) | SlowFast · Charades | mAP: 42.1 -> 42.5 | Improved |
| pre-train on Kinetics-600 instead of Kinetics-400 | SlowFast +NL · Charades | mAP: 42.5 -> 45.2 | Improved |
Closest papers by meaning
- SpeedNet: Learning the Speediness in Videos (CVPR 2020 · similarity 0.91)
- More Is Less: Learning Efficient Video Representations by Big-Little Network and Depthwise Temporal Aggregation (NeurIPS 2019 · similarity 0.88)
- FASTER Recurrent Networks for Efficient Video Classification (AAAI 2020 · similarity 0.88)
- Seeing Fast and Slow: Learning the Flow of Time in Videos (ECCV 2026 · similarity 0.88)
- Slow and Steady Feature Analysis: Higher Order Temporal Coherence in Video (CVPR 2016 · similarity 0.88)
- Action Recognition With Spatial-Temporal Discriminative Filter Banks (ICCV 2019 · 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.
