Research map / Video / Action recognition
FASTER Recurrent Networks for Efficient Video Classification
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 |
|---|---|---|---|
| aggregate mixed expensive and lightweight clip features with FAST-GRU instead of average pooling | FASTER · Kinetics val, 8-frame clips, 32 clips | top-1 accuracy: 63.8 -> 67.2 | Helped |
| use FAST-GRU instead of LSTM for aggregation | FASTER · Kinetics val, 8-frame clips, 32 clips | top-1 accuracy: 64.7 -> 67.2 | Helped |
| use FAST-GRU instead of GRU for aggregation | FASTER · Kinetics val, 8-frame clips, 32 clips | top-1 accuracy: 65.8 -> 67.2 | Helped |
| increase cheap clips while keeping only one expensive clip (1:7 rather than 1:1) (ablation) | FASTER · Kinetics val, 32-frame clips, 8 clips | top-1 accuracy: 74.6 -> 71.3 | Hurt |
| use fewer clips with longer clips at the same expensive-to-cheap ratio (32-frame clips rather than 8-frame clips, ratio 1:3) (ablation) | FASTER · Kinetics val, 1:3 expensive-to-cheap | top-1 accuracy: 70.1 -> 73.3 | Helped |
| replace the expensive R(2+1)D-50 backbone with R3D-50 (ablation) | FASTER32 · Kinetics val, 32-frame clips, 8 clips, 1:1 expensive-to-cheap | top-1 accuracy: 74.6 -> 75.3 | Helped |
| use FASTER32 with R3D-50 rather than average pooling over all expensive R3D-50 clips | FASTER32 · Kinetics val, 32-frame clips, 8 clips | top-1 accuracy: 74.4 -> 75.3 | Helped |
| use the hybrid FASTER32 setting rather than average pooling over all expensive clips | FASTER32 · Kinetics val, 32-frame clips, 8 clips | GFLOPs: 959.3 -> 550.4 | Helped |
Closest papers by meaning
- SpeedNet: Learning the Speediness in Videos (CVPR 2020 · similarity 0.89)
- Efficient Video Classification Using Fewer Frames (CVPR 2019 · similarity 0.88)
- Flexible Spatio-Temporal Networks for Video Prediction (CVPR 2017 · similarity 0.88)
- Recurrent Residual Module for Fast Inference in Videos (CVPR 2018 · similarity 0.87)
- Recurrent Dynamic Embedding for Video Object Segmentation (CVPR 2022 · similarity 0.87)
- FastVID: Dynamic Density Pruning for Fast Video Large Language Models (NeurIPS 2025 · similarity 0.86)
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.