Time and Memory Efficient 3D Point Cloud Classification

Recent advancement in real-time 3D scanning has upraised the need to maximally crunch the 3D data in minimum possible time. PointNet is a modern end-to-end deep learning architecture which directly utilizes point cloud without transforming it to regular formats for object-classification, part segmentation and scene semantic parsing. In this paper, we have proposed a novel set of hyper-parameters for PointNet which significantly reduced the size of network while achieving the same accuracy in 3D object classification as of the original network. This can enable the PointNet to be deployed on less resourceful hardware. The modified PointNet utilizes 16.5 Million less parameters (weights) and 5.2 Million lesser memory units. Moreover, through ample experiments, we demonstrate that our proposed configuration is highly efficient in terms of training and testing time.