H3DNET: A Deep Learning Framework for Hierarchical 3D Object Classification

Abstract

Deep learning has received a lot of attention in the fields such as speech recognition and image classification because of the ability to learn multiple levels of features from raw data. However, 3D deep learning is relatively new but in high demand with their great research values. Current research and usage of deep learning for 3D data suffer from the limited ability to process large volumes of data as well as low performance, especially in increasing the number of classes in the image classification task. One of the open questions is whether an efficient as well as an accurate 3D Deep Learning model can be built with large-scale 3D data. In this thesis, we aim to design a hierarchical framework for 3D Deep Learning, called H3DNET, which can build a DL 3D model in a distributed and scalable manner. In the H3DNET framework, a learning problem is composed of two stages: divide and conquer. At the divide learning stage, a learning problem is divided into several smaller problems. At the conquer learning stage, an optimized solution is used to solve these smaller subproblems for a better learning performance. This involves training of models and optimizing them with refined division for a better performance. The inferencing can achieve the efficiency and high accuracy with fuzzy classification using such a two-step approach in a hierarchical manner. The H3DNET framework was implemented in TensorFlow which is capable of using GPU computations in parallel to build 3D neural network. We evaluated the H3DNET framework on a 3D object classification with MODELNET10 and MODELNET40 datasets to check the efficiency of the framework. The evaluation results verified that the H3DNET framework supports hierarchical 3D Deep Learning with 3D images in a scalable manner. The classification accuracy is higher than the state-of-the-art, VOXNET[7] and POINTNET.