Object Classification and Shape Retrieval Based on 3D Object Surface

Abstract

This thesis proposes a machine learning based object classification method to recognize the object category when a single-view point cloud of an object surface is given. In addition, given multiple single-view surface point clouds capturing the shape of the object, a registration method that models the complete object shape is presented. Existing methods for object classification that use point clouds mostly depend on hand-crafted features. However, it is disadvantageous to use hand-crafted features in two aspects. First, they cannot fully utilize the raw data. Second, they lack extensibility in that they often require other types of feature such as surface normal. In this thesis, in order to circumvent these weaknesses, automatically learned features from voxelized grids of surface point clouds are used for classification. To this end, a deep learning architecture with Convolutional Neural Network (CNN) as a building block is proposed. However, most deep learning architectures suffer from the overfitting problem where the model fits even the stochastic noise in the training data. To reduce overfitting of the deep architecture, two types of non-random weight initialization methods are adopted. One is the initialization with the encoding stacks of a pre-trained de-noising Convolutional Auto-Encoder (DCAE) and the other is the initialization with the cluster centers learned by k-means clustering on a set of input data patches. Quantitative analysis shows that the suggested initialization methods reduce overfitting to a certain degree. The neural network is further designed to make predictions on the orientations of the surfaces as well as the class labels. It is shown analytically and quantitatively that this additional task improves the classification result. For the modeling of multiple single-view point clouds, a registration method that utilizes the Signature of Histogram of Orientation (SHOT) descriptors for correspondence association and Random Sample Consensus (RANSAC) algorithm for rigid transformation estimation is suggested. It is shown that by using the proposed RANSAC method to estimate the rigid transformations between successive point clouds and fine-tuning them by the Iterative Closest Point (ICP) algorithm, reliable object shape retrieval can be performed.