| Tartalmi leírás: |
Terrestrial laser scanners (TLS) capture structured panoramas in which every pixel back-maps to a 3D point, so a 2D image segmentation can be lifted directly into 3D. This makes vision foundation models attractive as a training-free front-end for point-cloud labelling, but a TLS panorama can be rendered from different per-point channels and it is unclear which one an RGB-pretrained model interprets best. We compare an open-vocabulary pipeline (GroundingDINO for text-driven detection followed by SAM2 for mask generation) on photographic RGB and laser-return intensity panoramas of the same scans, and isolate the role of the equirectangular projection by re-projecting to low-distortion cubemap views. Across indoor and outdoor scans the results indicate that RGB combined with cubemap re-projection is the most reliable configuration, that intensity is a surprisingly viable fallback, and that projection distortion measurably handicaps the detector on the raw 360° panorama.
|
