Title

3D POINT CLOUD CLASSIFICATION WITH GANS: ACGAN AND VACWGAN-GP

Abstract

Abstract

With the developing technology and the power of sensors, 3D data has started to be used in almost every field. Point clouds detected with LIDAR sensors or obtained by sampling 3D meshes have begun to come to the fore in many areas from autonomous driving to data visualization, from generating new data and mesh to classifying detected 3D objects. Machine learning and deep learning techniques are widely used to make sense of this produced data and to implement various applications. In this work, we propose networks to predict the class to which the 3D point cloud belongs, with Auxiliary Classifier Generative Adversarial Network and Versatile Auxiliary Conditional Wasserstein Generative Adversarial Network with Gradient Penalty, which are kind of GANs working with class labeled data. Unlike other classifiers; we are able to enlarge the limited data set with the data produced by taking advantage of the power of generative models, thus we aim to increase the success of the model by training it with more data. As suggested by the ACGAN models, the Discriminator is trained with synthetic data generated by the Generator with using the class label, in addition to the real dataset, ensures that data can be classified while separating real and fake data. Thus, as the training evolves, the Generator is trained to produce more realistic data; which forces Discriminator to classify better. Wasserstein GAN with GP demonstrates similar abilities with better training by replacing its Discriminator with Critic and modifying its loss function. In this work, we focus on merging Wassterstein GAN-GP with conditional GAN in order to improve the classifier’s performance. With this study, the proposed models were tested on 3D datasets and the results were compared with other studies.

Supervisor(s)

Supervisor(s)

ONUR ERGUN

Date and Location

Date and Location

2022-02-11 09:00:00

Category

Category

MSc_Thesis