Open-Vocabulary Object Detection on NWPU (test)
29.39mAP (PL)Geo-R1
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Geo-R1Optimization=DAPO, Protocol=10-shot Fine-tune, Samples=402025.09 | 29.39 | 24.49 | 68.2 | 17.04 | 34.78 | |
| Geo-R1Optimization=GRPO, Protocol=10-shot Fine-tune, Samples=402025.09 | 25.76 | 28.12 | 69.24 | 16.57 | 34.92 | |
| Qwen2.5-VLReasoning=w/ thinking, Protocol=Zero-shot Baseline2025.09 | 25.17 | 21.85 | 57.08 | 23.95 | 32.01 | |
| Qwen2.5-VLReasoning=w/o thinking, Protocol=Zero-shot Baseline2025.09 | 23.79 | 25.34 | 44.13 | 24.04 | 29.33 | |
| Geo-R1Optimization=DAPO, Protocol=5-shot Fine-tune, Samples=202025.09 | 22.15 | 25.57 | 72.12 | 14.7 | 33.64 | |
| Geo-R1Optimization=GRPO, Protocol=5-shot Fine-tune, Samples=202025.09 | 21.74 | 25.42 | 70.23 | 15.4 | 33.2 | |
| Qwen2.5-VL-SFTProtocol=10-shot Fine-tune, Samples=402025.09 | 15.76 | 21.9 | 68.42 | 14.73 | 30.2 | |
| Qwen2.5-VL-SFTProtocol=5-shot Fine-tune, Samples=202025.09 | 6.32 | 22.33 | 65.48 | 12.36 | 26.62 |