V2X-MGHD: A Collaborative Perception Network for Multiview LiDAR Sensors via Masked Generative Heterogeneous Distillation

Collaborative object detection in real-world scenarios often encounters issues of missed detections and false positives due to incomplete and sparse point cloud data. To address these challenges, we propose V2X-MGHD, a multiview LiDAR collaborative perception network based on a masked generative heterogeneous distillation framework. V2XMGHD offers three key advantages: First, to mitigate the problem of incomplete vehicle-mounted point cloud data, we introduce a mask generation heterogeneous distillation network that enhances the network’s capability for feature recovery and robust representation under feature-missing conditions. Additionally, to improve feature extraction efficiency and address the sparsity of large-scale environment point clouds, we propose the Sparse Inception Convolution (SIConv), which combines sparse convolution with Inception-style depthwise separable large kernel convolution. Lastly, to enhance robustness under communication delays and noisy environments, we incorporate a Temporal Synchronous Self-Attention (TSSA) module that extends multi-head self-attention with temporal position encoding, facilitating better synchronization and feature fusion over time. Comprehensive evaluations on the V2XSet and DAIR-V2X datasets demonstrate that V2X-MGHD outperforms the state-of-the-art method CoPEFT by 4.6% and 3.9% in detection accuracy, respectively. Moreover, V2X-MGHD significantly reduces inference time by 86%, showcasing its substantial practical value. The code for this work has been open-sourced and is available at: https://github.com/feeling0414-lab/V2X-MGHD.