Research map / Video / Action recognition
MARS: Motion-Augmented RGB Stream 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 |
|---|---|---|---|
| combine Flow-feature MSE with action cross-entropy in one RGB stream | 3D ResNeXt-101 · Kinetics400 | top-1 accuracy: 65.6 -> 72.7 | Helped |
| combine Flow-feature MSE with action cross-entropy in one RGB stream | 3D ResNeXt-101 · HMDB51-1 | top-1 accuracy: 73.5 -> 80.1 | Helped |
| combine Flow-feature MSE with action cross-entropy in one RGB stream | 3D ResNeXt-101 · SomethingSomethingv1 | top-1 accuracy: 43.0 -> 48.7 | Helped |
| replace explicit Flow input with RGB and train the stream to mimic Flow features (ablation) | MERS · MiniKinetics | top-1 accuracy: 61.7 -> 62.2 | No clear effect |
| add action cross-entropy training to the MSE-only feature-mimicking strategy (ablation) | MERS · MiniKinetics | top-1 accuracy: 62.2 -> 72.3 | Helped |
| set the feature-loss weight α to 50 (ablation) | MARS · MiniKinetics | top-1 accuracy | Helped |
| apply feature-based loss to earlier layers rather than high-level features (ablation) | MARS · MiniKinetics | top-1 accuracy | Hurt |
| apply feature-based loss to logits rather than high-level features (ablation) | MARS · MiniKinetics | top-1 accuracy | No clear effect |
| use 64-frame clips instead of 16-frame clips (ablation) | MARS · Kinetics400 | top-1 accuracy: 65.2 -> 72.7 | Helped |
| remove motion information by duplicating the middle frame (ablation) | MARS · MiniKinetics | top-1 accuracy: 72.3 -> 36.87 | Hurt |
| remove motion information by duplicating the middle frame (ablation) | MERS · MiniKinetics | top-1 accuracy: 62.2 -> 0.5% | Hurt |
Closest papers by meaning
- Scene Flow to Action Map: A New Representation for RGB-D Based Action Recognition With Convolutional Neural Networks (CVPR 2017 · similarity 0.90)
- Dance With Flow: Two-In-One Stream Action Detection (CVPR 2019 · similarity 0.90)
- Optical Flow Guided Feature: A Fast and Robust Motion Representation for Video Action Recognition (CVPR 2018 · similarity 0.89)
- Real-Time Action Recognition With Enhanced Motion Vector CNNs (CVPR 2016 · similarity 0.89)
- A Multi-Stream Bi-Directional Recurrent Neural Network for Fine-Grained Action Detection (CVPR 2016 · similarity 0.89)
- TrAction: Action Recognition with Sparse Trajectories (NeurIPS 2026 · 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.