Unsupervised Domain Adaptation for LiDAR Panoptic Segmentation

Scene understanding is a pivotal task for autonomous vehicles to safely navigate in the environment. Recent advances deep learning enable accurate semantic reconstruction of surroundings from LiDAR data. However, these models encounter large domain gap while deploying them on equipped with different setups which drastically decreases their performance. Fine-tuning model every new setup infeasible due expensive and cumbersome process recording manually labeling Unsupervised Domain Adaptation (UDA) techniques are thus essential fill this retain performance sensor without need additional data labeling. In paper, we propose AdaptLPS, novel UDA approach panoptic segmentation that leverages task-specific knowledge accounts variation number scan lines, mounting position, intensity distribution, environmental conditions. We tackle by employing two complementary adaptation strategies, data-based model-based. While adaptations reduce processing raw scans resemble target domain, model-based guide network extracting features representative both domains. Extensive evaluations three pairs real-world driving datasets demonstrate AdaptLPS outperforms existing approaches up 6.41 pp terms PQ score.