From sparse semantics to rich instances: Empowering label-efficient LiDAR panoptic segmentation via geometric priors

LiDAR point cloud panoptic segmentation is a rapidly developing task that unifies object detection, semantic segmentation, and instance segmentation. However, it requires both semantic and instance annotations, which are costly and labor-intensive to obtain. In this work, we propose a novel framework for LiDAR point cloud panoptic segmentation, which requires only a tiny proportion (e.g., 1% or fewer) of semantic labels and eliminates the need for any manual instance label. Specifically, we first adopt an active labeling strategy to annotate a small subset of the semantic labels. Then, we design a two-phase pipeline to generate and refine instance labels. In the Instance Generation Phase (IG-Phase), we compute category-aware instance priors based on the geometric characteristics of real-world LiDAR scenes and propose a heuristic clustering algorithm to automatically produce initial instance labels. In the Instance Refinement Phase (IR-Phase), these labels are further refined and corrected using semantic segmentation predictions and category-aware priors. In addition, we introduce a geometry-guided contrastive prototype learning module to enhance spatial feature aggregation and improve object discrimination by enforcing local semantic consistency. Extensive experiments on various large-scale LiDAR datasets with diverse backbones demonstrate the effectiveness of our approach. Under the same annotation budget, our framework remarkably outperforms traditional weakly-supervised annotations used in semantic segmentation and instance segmentation.