A deep learning based approach for saliency determination on point clouds

Laser scanners recording a huge number of data points from different surfaces are widely used to capture the exact geometry of objects. These large amounts of data require intelligent solutions to be examined and processed efficiently. Deep learning based approaches have found their way into many data analytic applications to process such large datasets, categorize them, or even determine the most informative portion of the data. This research focuses on 3D deep learning techniques directly applied to point clouds to determine the most important features of a 3D shape. More specifically this research adopts Pointnet as a backbone architecture for feature extraction from 3D point clouds and computes a Gradient-Based Class Activation Mapping on each object to create a 3D importance map for each object. Experiments confirm the success of the proposed approach in determination of important features of 3D objects as compared with the ground truth.