LiDAR point clouds are sparse, lack of feature and noisy. This makes perception and scene understanding extremely difficult with LiDAR data source. Here we proposed a novel method to tackle the sparsity 3D point cloud and achieve SOTA results on semantic KITTI scene completion benchmark.

We build our model based on Minkowski Engine. Since the Semantic Scene Completion data has uniformed voxel grid input (compared to raw point cloud), this property mitigates the affect of density difference of the LiDAR point cloud, thus we can safely apply the uniform receptive field to extract regional features in the 3D space.