PUPS: Point Cloud Unified Panoptic Segmentation

Point cloud panoptic segmentation is a challenging task that seeks holistic solution for both semantic and instance to predict groupings of coherent points. Previous approaches treat as surrogate tasks, they either use clustering methods or bounding boxes gather with costly computation hand-craft designs in the task. In this paper, we propose simple but effective point unified (PUPS) framework, which set point-level classifiers directly an end-to-end manner. To realize PUPS, introduce bipartite matching our training pipeline so are able exclusively instances, getting rid hand-crafted designs, e.g. anchors Non-Maximum Suppression (NMS). order achieve better grouping results, utilize transformer decoder iteratively refine develop context-aware CutMix augmentation overcome class imbalance problem. As result, PUPS achieves 1st place on leader board SemanticKITTI state-of-the-art results nuScenes.