SP2Mask4D: Efficient 4D Panoptic Segmentation Using Superpoint Transformers

The increasing need for precise segmentation in dynamic outdoor environments, particularly with LiDAR data, has brought attention to the 4D panoptic segmentation task. This task requires accurate identification of both objects and semantic labels across spatial and temporal dimensions. In this work, we present SP2Mask4D, a novel approach that replaces the commonly used transformer architecture with a superpoint-based transformer architecture. This modification leads to faster inference and reduced memory consumption, while maintaining competitive performance compared to transformer-based methods. While both approaches use attention mechanisms, traditional transformer models apply attention to all points, resulting in high computational costs. In contrast, SP2Mask4D focuses attention within localized superpoints, significantly lowering the computational burden. Experiments on the SemanticKITTI dataset show that SP2Mask4D reduces inference time by about 32.8% and improves memory efficiency by 60.3%, while preserving segmentation performance comparable to state-of-the-art methods.