Federated Grouping Panoptic Segmentation for LiDAR Point Cloud Based on Convolutional Network
Chengzhang Shi, Chung‐Ming Own, Ruimin Zhou · 2022
Panoptic segmentation of LiDAR point cloud in autonomous driving as a complete perception task has large data demand and does not facilitate data sharing when seeking collaboration. As a distributed collaborative learning method, Federated Learning (FL) is playing an increasingly important role in several fields. This paper is the first attempt to apply FL to LiDAR panoptic segmentation task. We propose a novel Panoptic Convolutional Fusion Network (PCF-Net), which outputs instance-level panoptic predictions directly through a learnable convolutional module. Additionally, we design the Dynamic Grouping Federated Learning (DGFL) algorithm, which transfers parameters based on grouping before global average. It maximizes the sharing times of model parameters while ensuring communication efficiency. Combining these two approaches, we propose a DGFL-based LiDAR panoptic segmentation framework. To demonstrate the effectiveness and generalizability of the framework, we reorganize and split two open-source datasets, namely SemanticKITTI and nuScenes. We conduct experiments on them under various conditions, and our method archives state-of-the-art performance. Our work provides new insights into the application of FL in autonomous driving.