Abstract
Accurate 3D semantic occupancy perception is critical for autonomous driving, enabling robust navigation in unstructured environments. While vision-based methods suffer from depth inaccuracies and lighting sensitivity, LiDAR-based approaches face challenges due to sparse data and dependence on expensive manual annotations. This work proposes LiGaussOcc, a novel self-supervised framework for dense LiDAR-based 3D semantic occupancy prediction. Our method first encodes LiDAR point clouds into voxel features and addresses sparsity via an Empty Voxel Inpainting (EVI) module, refined by an Adaptive Feature Fusion (AFF) module. During training, a Gaussian Primitive from Voxels (GPV) module generates parameters for 3D Gaussian Splatting, enabling efficient rendering of 2D depth and semantic maps. Supervision is achieved through photometric consistency across adjacent camera views and pseudo-labels from vision-language models, eliminating manual 3D annotations. Evaluated on the nuScenes-OpenOccupancy benchmark, LiGaussOcc achieved performance competitive with 30.4% Intersection over Union (IoU) and 14.1% mean Intersection over Union (mIoU). It reached 91.6% of the performance of the fully supervised LiDAR-based L-CONet, while completely eliminating the need for costly and labor-intensive manual 3D annotations. It excelled particularly in static environmental classes, such as drivable surfaces and man-made structures. This work presents a scalable, annotation-free solution for LiDAR-based 3D semantic occupancy perception.