上海 AI Lab 提出 OccAnyScene：一个模型统一室内外 3D 语义占据预测

TL;DR - OccAnyScene is a unified 3D semantic occupancy model that handles indoor rooms and outdoor roads despite differing cameras, spatial scales, voxel grids, and label systems. Joint training nearly matches scene-specific models, suggesting a path toward general-purpose occupancy foundation models.

• Uses pixel frustums as scale-adaptive geometric units and decodes them into 3D Gaussians representing visible and occluded regions.
• Shares all model parameters across indoor and outdoor scenes except dataset-specific semantic mapping matrices.
• The jointly trained model reaches 59.51% mIoU on Occ-ScanNet and 22.87% on SurroundOcc-nuScenes, within 0.41 and 0.19 points of scene-specific versions.
• Evaluation covers only two benchmarks, and areas outside camera-frustum coverage still require supplementary spatial queries.