Research map / Architectures and efficiency / CNN design
Representation Flow for Action Recognition
Check 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 |
|---|---|---|---|
| Compute representation flow on intermediate CNN features rather than RGB input (ablation) | Representation-flow ResNet-34 · Tiny-Kinetics | accuracy: 37.4 -> 59.4 | Helped |
| Compute representation flow after Block 3 rather than Block 5 (ablation) | Representation-flow ResNet-34 · Tiny-Kinetics | accuracy: 50.3 -> 59.4 | Helped |
| Learn divergence weights (ablation) | Representation-flow ResNet-34 · Tiny-Kinetics | accuracy: 59.4 -> 60.2 | Helped |
| Learn τ, λ, and θ (ablation) | Representation-flow ResNet-34 · Tiny-Kinetics | accuracy: 59.4 -> 59.9 | Helped |
| Learn Sobel kernel values (ablation) | Representation-flow ResNet-34 · Tiny-Kinetics | accuracy: 59.4 -> 58.5 | Hurt |
| Learn divergence weights together with τ, λ, and θ (ablation) | Representation-flow ResNet-34 · Tiny-Kinetics | accuracy: 59.4 -> 60.7 | Helped |
| Use 10 learned flow iterations rather than 1 (ablation) | Representation-flow ResNet-34 · Tiny-Kinetics | accuracy: 49.5 -> 59.4 | Helped |
| Add a convolution between two representation-flow layers (ablation) | Representation-flow ResNet-34 · Tiny-Kinetics | accuracy: 47.2 -> 62.3 | Helped |
| Add a third representation-flow layer (ablation) | Flow-Conv-Flow · Tiny-Kinetics | accuracy: 62.3 -> 56.5 | Hurt |
| Add representation flow to 3D CNN features after Block 3 (ablation) | 3D ResNet-18 · Tiny-Kinetics | accuracy: 54.6 -> 59.7 | Helped |
| Add representation flow to (2+1)D CNN features after Block 3 (ablation) | (2+1)D ResNet-18 · Tiny-Kinetics | accuracy: 53.4 -> 60.7 | Helped |
| Add a representation-flow layer to the 2D CNN | 2D CNN · Kinetics | accuracy: 61.3 -> 68.5 | Helped |
Closest papers by meaning
- Optical Flow Guided Feature: A Fast and Robust Motion Representation for Video Action Recognition (CVPR 2018 · similarity 0.91)
- Real-Time Action Recognition With Enhanced Motion Vector CNNs (CVPR 2016 · similarity 0.89)
- Scene Flow to Action Map: A New Representation for RGB-D Based Action Recognition With Convolutional Neural Networks (CVPR 2017 · similarity 0.88)
- Recognize Actions by Disentangling Components of Dynamics (CVPR 2018 · similarity 0.88)
- Deep Feature Flow for Video Recognition (CVPR 2017 · similarity 0.88)
- FlowFeat: Pixel-Dense Embedding of Motion Profiles (NeurIPS 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.