Research map / Video / Action recognition
Language-Driven Temporal Activity Localization: A Semantic Matching Reinforcement Learning Model
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 |
|---|---|---|---|
| Add semantic concept features to recurrent sentence-video matching (ablation) | Recurrent reinforcement learning localization · TACoS and Charades-STA | mR: 25.47 -> 28.24; 26.53 -> 31.45 | Helped |
| Use Faster R-CNN instead of multi-label classification to learn semantic concepts (ablation) | SM-RL with semantic concept features · TACoS and Charades-STA | mR: 26.80 -> 28.24; 28.49 -> 31.45 | Helped |
| Use frame-level classification loss instead of L2 boundary regression (ablation) | SM-RL with Faster R-CNN semantic concepts · TACoS and Charades-STA | mR: 28.24 -> 29.84; 31.45 -> 32.22 | Helped |
| Use sentence-video matching without semantic concepts (ablation) | Recurrent reinforcement learning localization · TACoS and Charades-STA | mR: 25.47 vs 27.80; 26.53 vs 30.24 | Hurt |
| Increase observations from 1 to 6 (ablation) | SM-RL(attr*+f) · TACoS | R@1, IoU=0.5: 4.26 -> 15.95 | Helped |
| Increase observations from 6 to 8 (ablation) | SM-RL(attr*+f) · TACoS | R@1, IoU=0.5: 15.95 -> 15.20 | Hurt |
| Set observation-location variance δ to 0.08 (ablation) | SM-RL(attr*+f) · TACoS | R@1, IoU=0.5: 9.57 -> 15.95; 9.36 -> 15.95; 10.16 -> 15.95; 8.93 -> 15.95 | Helped |
| Set loss weights to w1=1 and w2=1 (ablation) | SM-RL(attr*+f) · TACoS | R@1, IoU=0.5: 15.95 | Helped |
| Replace dense proposal evaluation with sequential selective frame observation | SM-RL · Video localization speed | average running time per minute video: 202ms -> 32ms | Helped |
Closest papers by meaning
- Semantic Proposal for Activity Localization in Videos via Sentence Query (AAAI 2019 · similarity 0.91)
- Read, Watch, and Move: Reinforcement Learning for Temporally Grounding Natural Language Descriptions in Videos (AAAI 2019 · similarity 0.88)
- Video-STAR: Reinforcing Open-Vocabulary Action Recognition with Tools (ICLR 2026 · similarity 0.88)
- Weakly Supervised Temporal Action Localization via Dual-Prior Collaborative Learning Guided by Multimodal Large Language Models (CVPR 2025 · similarity 0.87)
- Masked Diffusion Vision-Language Models for Temporal Action Localization (NeurIPS 2026 · similarity 0.87)
- VideoGEM: Training-free Action Grounding in Videos (CVPR 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.