Deep Learning Based Segmentation of 3D Point Clouds

Abstract

3D models like point clouds are very widespread these days and are being used in numerous fields. Despite their broad availability, there is still a perti nent need of automatic approaches for semantic segmentation of data in 3D. Semantic segmentation is an important job and is a primary step towards scene understanding. Segmentation of 3D point clouds is a difficult task es pecially handling large scale point clouds is still an open challenge. Many deep learning architectures have been designed which can take unordered point clouds as input but a very few addressed the problem of segmenting large scale point clouds. Hence it is needed to design a segmentation archi tecture which is efficient enough to handle point-clouds in large scale while keeping model fast and accurate. Recently, Landrieu et al. [24] proposed an impressive model for large scale point set segmentation. The model showed remarkable performance but used fixed adjacency super point graphs for vari able density point-clouds, decreasing their embedding quality. Hence efficient partitioning of large point-clouds is needed, while keeping local geometric re lationships in mind. In this regard, we have proposed a model for semantic segmentation of point clouds in large scale, which is a decent combination of unsupervised and supervised machine learning approaches. The model works on eloquent partitioning of point-clouds, which are based on local geometric dependencies between points. Our model not only outperforms other state of the art architectures in time and memory but also has shown comparable accuracy with other successful architectures.