Research map / Video / Video-language understanding
Read, Watch, and Move: Reinforcement Learning for Temporally Grounding Natural Language Descriptions in Videos
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 |
|---|---|---|---|
| Use sequential RL to iteratively adjust temporal boundaries rather than exhaustive candidate ranking or sliding windows | Video grounding · ActivityNet DenseCaption and Charades-STA | Acc@0.5: 36.9% vs. 34.7%; 36.7% vs. 32.5% | Helped |
| Add explicit temporal location features to the observation (ablation) | Ours · ActivityNet DenseCaption and Charades-STA | Acc@0.5: 33.5% -> 36.9%; 34.8% -> 36.7% | Helped |
| Add location regression supervision (ablation) | Ours · ActivityNet DenseCaption and Charades-STA | Acc@0.5: 35.9% -> 36.9%; 35.1% -> 36.7% | Helped |
| Add tIoU regression supervision alongside location regression (ablation) | Ours · ActivityNet DenseCaption and Charades-STA | Acc@0.5: 34.5% -> 36.9%; 34.9% -> 36.7% | Helped |
| Add all three components: explicit location feature, tIoU regression, and location regression (ablation) | Ours · ActivityNet DenseCaption and Charades-STA | Acc@0.5: 31.5% -> 36.9%; 13.0% -> 36.7% | Helped |
| Reduce step penalty φ from 1 to 0.001 (ablation) | Ours · ActivityNet DenseCaption | Acc@0.5: 25.4% -> 36.9% | Helped |
| Reduce step penalty φ from 1 to 0.001 (ablation) | Ours · Charades-STA | Acc@0.5: 17.9% -> 36.7% | Helped |
| Reduce step penalty φ from 1 to 0.001 (ablation) | Ours · ActivityNet DenseCaption | # Steps: 1.00 -> 9.99 | Hurt |
| Reduce step penalty φ from 1 to 0.001 (ablation) | Ours · Charades-STA | # Steps: 1.0 -> 9.99 | Hurt |
| Use an intermediate discount factor instead of γ=0 or γ=1 (ablation) | Ours · ActivityNet DenseCaption and Charades-STA | Acc@0.5 | Helped |
| Fuse RGB and optical-flow features by concatenation instead of using either modality alone (ablation) | Ours · Charades-STA | Acc@0.5: 36.2% and 35.8% -> 36.7% | Helped |
Closest papers by meaning
- Video-OPD: Efficient Post-Training of Multimodal Large Language Models for Temporal Video Grounding via On-Policy Distillation (ICML 2026 · similarity 0.91)
- Uncertainty-quantified Rollout Policy Adaptation for Unlabelled Cross-domain Video Temporal Grounding (NeurIPS 2025 · similarity 0.90)
- Tree-Structured Policy Based Progressive Reinforcement Learning for Temporally Language Grounding in Video (AAAI 2020 · similarity 0.89)
- EvoGround: Self-Evolving Video Agents for Video Temporal Grounding (NeurIPS 2026 · similarity 0.89)
- Time-R1: Post-Training Large Vision Language Model for Temporal Video Grounding (NeurIPS 2025 · similarity 0.88)
- EventLens: Event-Structure Reinforcement Learning for Video Understanding (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.