Bounding box detection on RefCOCOg (val)
81.01mIoU (bbox)GenSeg-R1-4B
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| GenSeg-R1-4BBase VLM=Qwen3-VL-4B, SAM 2 model=sam2-hiera-large, Training protocol=RL-based2026.02 | 81.01 | 73.47 | 87.8 | 78.9 | |
| GenSeg-R1-8BBase VLM=Qwen3-VL-8B, SAM 2 model=sam2-hiera-large, Training protocol=RL-based2026.02 | 80.28 | 72.77 | 87.17 | 78.35 | |
| Seg-Zero-7BBase VLM=Qwen2.5-VL-7B, SAM 2 model=sam2-hiera-large, Training protocol=RL-based2026.02 | 80 | 71.97 | — | — | |
| VisionReasoner-7BBase VLM=Qwen2.5-VL-7B, SAM 2 model=sam2-hiera-large2026.02 | 79.67 | 71.49 | 85.81 | 76.02 | |
| Seg-R1-7BBase VLM=Qwen2.5-VL-7B, SAM 2 model=sam2-hiera-large, Training protocol=RL-based2026.02 | 74.52 | 62.94 | 82.16 | 70.15 | |
| Qwen2.5-VL-7B-Instruct2026.02 | 58.26 | 51.62 | 62.73 | 55.15 | |
| Qwen3-VL-8B-Instruct2026.02 | 55.46 | 49.81 | 59.93 | 53.25 | |
| Qwen3-VL-4B-Instruct2026.02 | 53.01 | 47.45 | 57.02 | 50.49 |