Bootstrapping a 4D LiDAR Annotation Tool from Video Foundation Models
LiDARSAM2 uses a 2D video foundation model (SAM2) to automatically generate scalable supervision for 4D LiDAR segmentation, overcoming data scarcity. The framework treats SAM2 as a source of pseudo-labels, using multi-view projections and spatio-temporal aggregation to create consistent 4D masks. This reduces the need for manual labeling while achieving high quality in scene understanding.