Fine localization on LangLoc (full (1,300 scenes))
1.369Positional Error (m) MeanMidpoint
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| MidpointLogic=floor centroid (no language)2026.07 | 1.369 | — | — | 1.259 | — | — | — | |
| LangLoc w/o dialogDialog=without dialog2026.07 | 1.534 | 1.153 | 0.951 | 1.308 | 46.85 | 39.8 | 14.7 | |
| VLM (Qwen2.5-VL)Model=Qwen2.5-VL-2B, Rendering=top-down, Mode=zero-shot2026.07 | 1.538 | — | — | 1.35 | 84.7 | 81.7 | 1.8 |