STMOS: Spatio-Temporal Transformer for Moving Object Segmentation

We present STMOS, a novel spatio-temporal transformer architecture for moving object segmentation in LiDAR point clouds. While recent approaches using sparse 4D convolutions have shown promising results for this task, they often struggle to capture long-range dependencies crucial for understanding complex motion patterns. STMOS addresses this limitation by integrating transformer components with sparse 4D convolutions, creating a hybrid architecture that effectively combines local feature extraction with global context modeling. Our model employs a specially designed sparse attention mechanism that considers spatial proximity in point clouds, significantly reducing computational costs while maintaining the ability to model relationships between distant points. Extensive experiments on the SemanticKITTI benchmark demonstrate that STMOS outperforms state-of-the-art methods, particularly in challenging scenarios involving partially occluded objects and complex motion patterns. The hybrid nature of our architecture ensures computational efficiency comparable to pure convolutional approaches, making it suitable for real-time autonomous driving applications.