A voxel-based deep learning approach for Point Cloud Semantic Segmentation

Semantic segmentation has been a research topic in computer vision for decades. This task has become a crucial challenge nowadays due to emergence of new technologies such as autonomous driving. Nonetheless, most existing segmentation methods are not designed for handling the unstructured and irregular nature of 3D point clouds. We propose a voxel-based technique for 3D point cloud data semantic segmentation using 3D convolutional neural networks. It uses local voxelizations for learning spatial patterns, and also corrects the imbalance of the data, something very problematic with 3D datasets.