Real-world cross-view spatial reasoning on STARE
35.6AccuracyPanoramic + VDrop
Evaluation Results
| Method | Links | |
|---|---|---|
| Panoramic + VDropVDrop=✓2026.05 | 35.6 | |
| Point Matching + VDropVDrop=✓2026.05 | 34.4 | |
| PanoramicVDrop=✗2026.05 | 32.4 | |
| Qwen3-VL-8BVDrop=✗2026.05 | 31.6 | |
| Qwen3-VL-4BVDrop=✗2026.05 | 29.2 | |
| ThinkMorph (24K)VDrop=✗2026.05 | 28.8 | |
| BAGEL-Zebra-CoT (182K)VDrop=✗2026.05 | 28.4 | |
| BAGELVDrop=✗2026.05 | 28 | |
| Top-downVDrop=✗2026.05 | 28 | |
| Text CoTVDrop=✗2026.05 | 26 | |
| Top-down + VDropVDrop=✓2026.05 | 26 | |
| No-ThinkVDrop=✗2026.05 | 24.8 | |
| Point MatchingVDrop=✗2026.05 | 24 |