Sa2VA: Marrying SAM2 with LLaVA for Dense Grounded Understanding of Images and Videos

Benchmark Model Rank Results
referring-expression-segmentation-on-refer-youtube-vos-2021-public-validationSa2VA-26B (extra data)#1J&F: 75.1
referring-video-object-segmentation-on-ref-davis17Sa2VA-Qwen2.5VL-7B (extra data)#5J&F: 79.4
referring-video-object-segmentation-on-revosSa2VA-InternVL3-8B (extra data)#11J&F: 60.8
referring-video-object-segmentation-on-mevisSa2VA-8B (extra data)#23J&F: 46.9