Open-Vocabulary Detection on NWPU VHR-10 (val)
26mAP (IoU=0.5:0.95)RS-HyRe-R1
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| RS-HyRe-R1Base LLM=Qwen2.5-3B, Protocol=RL Train2026.04 | 26 | 52.25 | |
| Geo-R1-OVDBase LLM=Qwen2.5-3B, Protocol=RL Train2026.04 | 18.87 | 34.28 | |
| GeoChatBase LLM=Vicuna1.5-7B, Protocol=RS-task Dataset Fine-tune (1600 samples)2026.04 | 17.31 | 35.09 | |
| Qwen2.5-VL-SFTBase LLM=Qwen2.5-3B, Protocol=RS-task Dataset Fine-tune (1600 samples)2026.04 | 15.92 | 34.74 | |
| Geo-R1Base LLM=Qwen2.5-7B, Protocol=RL Train2026.04 | 15.61 | 33.58 | |
| GeoReasonBase LLM=Qwen2.5-7B, Protocol=RL Train2026.04 | 12.05 | 29.16 | |
| Geo-R1-RECBase LLM=Qwen2.5-3B, Protocol=RL Train2026.04 | 10.67 | 26.12 | |
| Qwen2.5-VL 7BBase LLM=Qwen2.5-7B, Protocol=Zero-shot Baseline2026.04 | 10.44 | 15.39 | |
| R1-VLBase LLM=Qwen2-7B, Protocol=RL Train2026.04 | 10.12 | 25.9 | |
| TinyRS-R1Base LLM=Qwen2-2B, Protocol=RL Train2026.04 | 8.73 | 16.05 | |
| VLM-R1-OVDBase LLM=Qwen2.5-3B, Protocol=RL Train2026.04 | 7.4 | 19.01 | |
| VLM-R1-RECBase LLM=Qwen2.5-3B, Protocol=RL Train2026.04 | 6.03 | 16.58 | |
| Qwen2.5-VL 3BBase LLM=Qwen2.5-3B, Protocol=Zero-shot Baseline2026.04 | 3.86 | 9.48 |