Proposing SAL for zero-shot Lidar segmentation and classification, distilling vision models into Lidar for universal applications.
SAL, a method for Zero-Shot Lidar Panoptic Segmentation, utilizes a pseudo-label engine that distills vision foundation models to Lidar and a zero-shot model trained via self-supervision. This allows for segmenting and classifying any object in a Lidar scan without manual supervision.