Research map / Video / Video-language understanding
Sa2VA: Marrying SAM2 With MLLM for Dense Grounded Understanding of Images and Videos
Haobo Yuan, Xiangtai Li, Tao Zhang, Yueyi Sun, Zilong Huang, Shilin Xu, Shunping Ji, Yunhai Tong, Lu Qi, Jiashi Feng, Ming–Hsuan Yang
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 image QA data to co-training (ablation) | Sa2VA-1B · MMBench / MME | accuracy: 63.4 -> 68.3 MMBench; 1298/359 -> 1381/405 MME | Improved |
| add image segmentation data to co-training (ablation) | Sa2VA-1B · RefCOCO / RefCOCO+ / RefCOCOg | cIoU: 20.2 -> 77.4; 20.6 -> 69.9; 23.2 -> 72.3 | Improved |
| add video QA data to co-training (ablation) | Sa2VA-1B · MMBench-Video | score: 0.71 -> 1.07 | Improved |
| add video segmentation data to co-training (ablation) | Sa2VA-1B · MeViS / Ref-DAVIS17 | J&F: 44.4 -> 50.8 MeViS; 69.0 -> 72.3 Ref-DAVIS17 | Improved |
| add 3M Infinity-MM image QA samples | Sa2VA-1B · MMBench | accuracy: 68.3 -> 70.4 | Improved |
| add Ref-SAV training data | Sa2VA-1B · MeViS (val_u) | J&F: 50.8 -> 52.5 | Improved |
| fine-tune UniRef++ on Ref-SAV training set | UniRef++ · Ref-SAV val | J&F: 10.5 -> 14.6 | Improved |
| sample uniform 5 frames instead of first 5 frames (ablation) | Sa2VA-8B · MeViS (val_u) | J&F: 58.9 -> 62.9 | Improved |
| use more consecutive first frames (1 to 4) (ablation) | Sa2VA-8B · MeViS (val_u) | J&F: 55.1 -> 59.5 | Improved |
| replace SAM-2 with SAM-3 segmenter (ablation) | Sa2VA (Qwen3-VL-4B) · MeViS / ReVOS | J&F: 61.4 -> 65.3 MeViS; 61.4 -> 66.3 ReVOS | Improved |
| use Sa2VA paradigm with LLaVA-1.5-7B instead of LISA-7B | LISA-7B · RefCOCO / RefCOCO+ / RefCOCOg | cIoU: 74.9 -> 80.3 RefCOCO; 65.1 -> 73.1 RefCOCO+; 67.9 -> 79.2 RefCOCOg | Improved |
| scale Sa2VA from 8B to 26B | Sa2VA · MeViS | J&F: 46.9 -> 46.2 | No clear effect |
Later papers that compared with this method
5 later accepted papers report Sa2VA-8B* as a baseline in their results tables; 5 of them report a better number on most of the results they share with it.
- RVAS: Referring Video Active Exploration and Segmentation (ICML 2026 · better on 7 of 8 shared results)
- VideoLoom: A Video Large Language Model for Joint Spatial-Temporal Understanding (ICML 2026 · better on 4 of 4 shared results)
- VideoSEG-O3: A Multi-turn Reinforcement Learning Framework for Reasoning Video Object Segmentation (ICML 2026 · better on 1 of 1 shared result)
- Open-o3-Video: Grounded Video Reasoning with Explicit Spatio-Temporal Evidence (ICML 2026 · better on 1 of 1 shared result)
- MomentSeg: Moment-Centric Sampling for Enhanced Referring Video Object Segmentation (ECCV 2026 · better on 1 of 1 shared result)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- VISA-13B — VISA: Reasoning Video Object Segmentation via Large Language Model (ECCV 2024) · better on 3 of 3 shared results
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
- ViCaS: A Dataset for Combining Holistic and Pixel-level Video Understanding using Captions with Grounded Segmentation (CVPR 2025 · similarity 0.89)
- SAMWISE: Infusing Wisdom in SAM2 for Text-Driven Video Segmentation (CVPR 2025 · similarity 0.89)
- SAM 2: Segment Anything in Images and Videos (ICLR 2025 · similarity 0.88)
- GSVA: Generalized Segmentation via Multimodal Large Language Models (CVPR 2024 · similarity 0.88)
- SAM-I2V: Upgrading SAM to Support Promptable Video Segmentation with Less than 0.2% Training Cost (CVPR 2025 · similarity 0.88)
- SAMA: Towards Multi-Turn Referential Grounded Video Chat with Large Language Models (NeurIPS 2025 · similarity 0.88)
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.
