Generalized Referring Expression Segmentation on EarthReason (test)
70.75gIoUSegEarth-R1
Evaluation Results
| Method | Links | |
|---|---|---|
| SegEarth-R1Evaluation Protocol=Full Amount Fine-tune2025.09 | 70.75 | |
| PSALMEvaluation Protocol=Full Amount Fine-tune2025.09 | 68.3 | |
| LISAEvaluation Protocol=Full Amount Fine-tune2025.09 | 60.88 | |
| PixelLMEvaluation Protocol=Full Amount Fine-tune2025.09 | 60.01 | |
| Geo-R1 (DAPO)Evaluation Protocol=10-shot Fine-tune, Number of training samples=2402025.09 | 58.41 | |
| Geo-R1 (GRPO)Evaluation Protocol=10-shot Fine-tune, Number of training samples=2402025.09 | 58.27 | |
| SegEarth-R1Evaluation Protocol=10-shot Fine-tune, Number of training samples=2402025.09 | 56.6 | |
| Geo-R1 (DAPO)Evaluation Protocol=5-shot Fine-tune, Number of training samples=1202025.09 | 56.24 | |
| Geo-R1 (GRPO)Evaluation Protocol=5-shot Fine-tune, Number of training samples=1202025.09 | 56.01 | |
| Geo-R1 (DAPO)Evaluation Protocol=1-shot Fine-tune, Number of training samples=242025.09 | 51.82 | |
| Geo-R1 (GRPO)Evaluation Protocol=1-shot Fine-tune, Number of training samples=242025.09 | 51.38 | |
| SegEarth-R1Evaluation Protocol=5-shot Fine-tune, Number of training samples=1202025.09 | 45.46 | |
| SegEarth-R1Evaluation Protocol=1-shot Fine-tune, Number of training samples=242025.09 | 43.01 | |
| Qwen2.5-VL w/ thinkingEvaluation Protocol=Zero-shot Baseline2025.09 | 32.16 |