GeoOpen3D: Geometry-guided training-free open-vocabulary 3D segmentation via visual foundation models

GeoOpen3D adapts visual foundation models to outdoor point-cloud segmentation without training a 3D network. Geometry-preserving RGB-D projections improve the input to GroundingDINO and SAM, while depth-aware constraints select coherent masks. Pixel-to-point mappings then transfer the resulting labels into 3D. Evaluations on urban and island scenes examine segmentation quality and transfer to categories outside a fixed label set.

The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, XLIX-B2-2026, pp. 1335–1343