ST-CFNet: A Spatio-Temporal Enhanced Network for Real-Time 4d Panoptic Segmentation

Real-time scene understanding is vital for autonomous driving, robotics, and augmented reality. 4D panoptic LiDAR segmentation extends 3D segmentation by adding temporal consistency of instance IDs over sequential point clouds. Recent transformer-based methods often fail to achieve real-time performance due to their high computational complexity. To address this, we propose ST-CFNet, a spatio-temporal extension of CFNet. It processes spatio-temporal point cloud to capture temporal and geometric features, employs a 2D feature projection backbone with the TWA-P2G module for efficient temporal feature aggregation, and generates predicted results with a 4D panoptic segmentation head. Finally, a center–box deduplication module is adopted to select the center for each detected instance. ST-CFNet achieves a score of 70.2 LSTQ on the SemanticKITTI validation set, outperforming most existing methods while enabling real-time 4D panoptic segmentation. The code is available at https://github.com/ny823/ST-CFNet.