Research map / Multi-agent systems and games / Cooperative multi-agent RL
Multi-Agent Reinforcement Learning Based Frame Sampling for Effective Untrimmed Video Recognition
Wenhao Wu, Dongliang He, Xiao Tan, Shifeng Chen, Shilei Wen
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 uniform 25-frame sampling with MARL sampling | ResNet-152 · ActivityNet v1.3 val (ResNet-152) | mAP: 82.08 -> 83.81 | Improved |
| replace uniform sampling with MARL sampling | ResNet-152 · ActivityNet v1.2 val (ResNet-152) | mAP: 83.71 -> 85.70 | Improved |
| replace uniform sampling with MARL sampling on C3D | C3D · ActivityNet v1.3 val (C3D) | mAP: 60.68 -> 62.00 | Improved |
| add context-aware observation (M=1 vs M=0) (ablation) | MARL (ResNet-152) · ActivityNet v1.3 val (ResNet-152) | mAP: 82.99 -> 83.81 | Improved |
| increase context range from M=1 to M=4 (ablation) | MARL (ResNet-152) · ActivityNet v1.3 val (ResNet-152) | mAP: 83.81 -> 83.72 | No clear effect |
| increase number of agents at test from 5 to 25 (ablation) | MARL (ResNet-152) · ActivityNet v1.3 val (ResNet-152) | mAP: 80.19 -> 83.81 | Improved |
| increase test agents from 25 to 120 (all frames) (ablation) | MARL (ResNet-152) · ActivityNet v1.3 val (ResNet-152) | mAP: 83.81 -> 83.72 | No clear effect |
| train with video-level labels only instead of instance-level (ablation) | MARL (ResNet-152) · ActivityNet v1.3 val (ResNet-152) | mAP: 83.81 -> 82.98 | Hurt |
| transfer policy network trained on YouTube Birds to ActivityNet v1.3 (ablation) | MARL (ResNet-152) · ActivityNet v1.3 val (ResNet-152) | mAP: 83.81 -> 82.70 | Mixed |
| transfer policy trained for ResNet-101 to ResNet-152 classifier (ablation) | MARL (ResNet-152) · ActivityNet v1.3 val (ResNet-152) | mAP: 83.81 -> 83.43 | Mixed |
| add MARL sampling to Inception-V3 on YouTube Birds vs RRA | RRA · YouTube Birds | top-1 accuracy: 73.205 -> 79.01 | Improved |
| add MARL sampling vs uniform sampling on YouTube Cars | Inception-V3 · YouTube Cars | top-1 accuracy: 76.49 -> 79.77 | Improved |
Closest papers by meaning
- Attention-Aware Sampling via Deep Reinforcement Learning for Action Recognition (AAAI 2019 · similarity 0.88)
- ReAgent-V: A Reward-Driven Multi-Agent Framework for Video Understanding (NeurIPS 2025 · similarity 0.87)
- Search-Map-Search: A Frame Selection Paradigm for Action Recognition (CVPR 2023 · similarity 0.86)
- Videos are Sample-Efficient Supervisions: Behavior Cloning from Videos via Latent Representations (NeurIPS 2025 · similarity 0.86)
- PyVision-RL: Forging Open Agentic Vision Models via RL (ICML 2026 · similarity 0.85)
- Towards Principled Representation Learning from Videos for Reinforcement Learning (ICLR 2024 · similarity 0.84)
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.
