Semantic Segmentation of 3D Point Clouds using Deep Learning

Abstract

This thesis presents a new fine-grained analysis on large-scale point cloud semantic segmentation through a mechanism which fuses a local and global context aware feature representation strategy into a semantic segmentation model. Real point cloud scenes can acquire intrinsic details of the object present in the scene, but due to raw structure of point cloud data, some im portant details can be lost in the process semantic segmentation. Moreover, effective feature learning for large-scale point clouds is inherently a complex task. In this thesis, spatial contextual features are learned locally and glob ally and then fused into a semantic segmentation model which augments the geometric and semantic features in the point cloud for the task of semantic segmentation of point clouds. Further, several multi-resolution point cloud scenes have been adaptively fused to leverage better semantic segmentation results. Benchmark dataset S3DIS of large scale point clouds have been used for training and evaluation.