Executive Summary : | The processing of 3D data using traditional methods is challenging due to the lack of standard deep learning techniques from other domains. This project aims to develop a method to learn qualitative 3D point cloud representations, considering scenarios such as unavailability of labeled data, limited labeled data availability, and limited resources for target distribution. Self-supervised learning (SSL) is proposed as a solution, which has proven effective for unlabeled data. Prototype learning (PL) and domain adaptation (DA) are popular methods for solving learning problems in unsupervised, semi-supervised, and self-supervised settings. The development of a 3D point cloud model will involve a thorough study on geometrical and structural aspects of 3D data. The idea of SSL for 3D scenarios will be explored with optimal transport (OT) to better understand geometry variations in the data. The project will focus on two sub-problems: the unlabeled and limited labeled data problem using PL and OT, and the domain adaptation problem using shared embedding for both domains while minimizing Wasserstein distance. Extensive analysis will be conducted on self-supervised, few-shot, prototype, cross-domain learning, and optimal transport for 3D point cloud representations. Technical research will be conducted in domain adaptation for 3D point clouds and metric learning in 3D space to build state-of-the-art geometry-aware 3D point cloud representation learning methods. The project's outcomes will significantly improve 3D deep learning methods and can be developed into an independent software/product for use in most 3D models across various domains. |