Research map / Video / Video-language understanding
VideoSEG-O3: A Multi-turn Reinforcement Learning Framework for Reasoning Video Object Segmentation
Ming Dai, Sen Yang, Boqiang Duan, Boyuan Tong, Jiedong Zhuang, Wankou Yang, Jingdong Wang
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 CoT cold-start stage on top of SFT (ablation) | VideoSEG-O3-4B · Long-RVOS | J&F: 47.12 -> 51.82 | Improved |
| Add CoT cold-start stage on top of SFT (ablation) | VideoSEG-O3-4B · Ref-SAV | J&F: 51.97 -> 50.11 | Hurt |
| Add RL stage after CoT cold-start (ablation) | VideoSEG-O3-4B · Six RVOS benchmarks average | Avg. J&F: 57.07 -> 60.29 | Improved |
| Add RL stage after CoT cold-start (ablation) | VideoSEG-O3-4B · Six RVOS benchmarks average | Avg. rounds: 2.79 -> 2.23 | Improved |
| Use standard GRPO with text-only importance ratio (ablation) | Stage II cold-start model · MeViS (valu) | J&F: 59.43 -> 57.92 | Hurt |
| Add SEG-aware logit calibration to GRPO (ablation) | GRPO-RL baseline · Ref-SAV | J&F: 49.78 -> 51.27 | Improved |
| Add SEG-aware logit calibration to GRPO (ablation) | GRPO-RL baseline · MeViS (valu) | J&F: 57.92 -> 60.51 | Improved |
| Add spatial-wise auxiliary segmentation loss (ablation) | GRPO + SEG calibration · MeViS (valu) | J&F: 60.51 -> 60.71 | Improved |
| Add keyframe-wise auxiliary segmentation loss (ablation) | GRPO + SEG calibration · MeViS (valu) | J&F: 60.51 -> 61.31 | Improved |
| Use joint spatial and keyframe auxiliary loss (ablation) | GRPO + SEG calibration · MeViS (valu) | J&F: 60.51 -> 61.95 | Improved |
| Use joint auxiliary segmentation loss (ablation) | GRPO + SEG calibration · Ref-SAV | J&F: 51.27 -> 53.22 | Improved |
| Apply RL to keyframe selection (ablation) | VideoSEG-O3-4B · four RVOS benchmarks | keyframe mIoU and empty ratio | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- Sa2VA-8B — Sa2VA: Marrying SAM2 With MLLM for Dense Grounded Understanding of Images and Videos (TPAMI 2026) · better on 1 of 1 shared result
- GLUS-7B — GLUS: Global-Local Reasoning Unified into A Single Large Language Model for Video Segmentation (CVPR 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- The Devil is in Temporal Token: High Quality Video Reasoning Segmentation (CVPR 2025 · similarity 0.91)
- Fast Reasoning Segmentation for Images and Videos (CVPR 2026 · similarity 0.90)
- Video-R1: Reinforcing Video Reasoning in MLLMs (NeurIPS 2025 · similarity 0.90)
- LENS: Learning to Segment Anything with Unified Reinforced Reasoning (AAAI 2026 · similarity 0.90)
- When Thinking Drifts: Evidential Grounding for Robust Video Reasoning (NeurIPS 2025 · similarity 0.89)
- Video-of-Thought: Step-by-Step Video Reasoning from Perception to Cognition (ICML 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.
