Research map / Video / Video-language understanding
SARL-STG: A Spatially Aware Reinforcement Learning Framework for Refining MLLMs in Spatio-Temporal Video Grounding
Hong Gao, Xiangkai Xu, Bin Zhong, Junjie Yin, Fangyu Kang, Yutong Xu, Xiugang Dong, Xurui Gao, Min-Ling Zhang
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 |
|---|---|---|---|
| add spatio-temporal dynamic reward in GRPO stage on top of joint SFT (ablation) | SARL-STG · HCSTVG | m tIoU: 58.6 -> 64.2 | Improved |
| use dynamic tIoU-weighted reward instead of plain spatio-temporal reward (ablation) | SARL-STG GRPO · VidSTG | m tIoU: 51.5 -> 52.3 | Improved |
| add spatial knowledge to reward (spatio-temporal reward vs temporal-only reward) (ablation) | SARL-STG GRPO · HCSTVG | m tIoU: 62.3 -> 63.8 | Improved |
| switch from temporal-only SFT to joint spatio-temporal SFT (ablation) | SARL-STG Stage 1 · HCSTVG | m tIoU: 62.5 -> 58.6 | Hurt |
| switch from temporal-only SFT to joint spatio-temporal SFT (ablation) | SARL-STG Stage 1 · VidSTG | m tIoU: 47.7 -> 50.5 | Improved |
| add GRPO with temporal reward after SFT (ablation) | SARL-STG · VidSTG | m tIoU: 50.5 -> 51.3 | Improved |
| add Temporal Enhancement Module and Global-Local Feature Fusion Module to spatial grounding (ablation) | GroundingDINO · HCSTVG, GT input | m sIoU: 65.6 -> 72.1 | Improved |
| add Temporal Enhancement Module alone (ablation) | GroundingDINO · HCSTVG, GT input | m sIoU: 65.6 -> 71.0 | Improved |
| shift input clip randomly by 40% from ground truth (ablation) | spatial grounding module · HCSTVG | m sIoU: 72.1 -> 66.4 | Hurt |
| shift input clip randomly by 20% from ground truth (ablation) | spatial grounding module · HCSTVG | m sIoU: 72.1 -> 70.3 | Hurt |
| train on STVG-Wild instead of STVG data only (ablation) | SARL-STG · ST-Align (OOD) | m tIoU: 42.6 -> 49.5 | Improved |
| add TVG data to STVG training data (ablation) | SARL-STG · ActivityNet (OOD) | m tIoU: 28.5 -> 33.4 | Improved |
Closest papers by meaning
- Unleashing the Potential of Multimodal LLMs for Zero-Shot Spatio-Temporal Video Grounding (NeurIPS 2025 · similarity 0.92)
- Time-R1: Post-Training Large Vision Language Model for Temporal Video Grounding (NeurIPS 2025 · similarity 0.91)
- STPro: Spatial and Temporal Progressive Learning for Weakly Supervised Spatio-Temporal Grounding (CVPR 2025 · similarity 0.91)
- Knowing Your Target: Target-Aware Transformer Makes Better Spatio-Temporal Video Grounding (ICLR 2025 · similarity 0.89)
- Grounded Reinforcement Learning for Visual Reasoning (NeurIPS 2025 · similarity 0.89)
- Context-Guided Spatio-Temporal Video Grounding (CVPR 2024 · similarity 0.89)
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.
