CMDGAT: Knowledge extraction and retention based continual graph attention network for point cloud registration.

Artificial Intelligence-based systems are required to interact with dynamic environments to continuously learn, retain and effectively utilize knowledge. Present AI-based systems scarcely exhibit this ability and mostly follow an isolated learning paradigm. In contrast, a human brain goes through a continual learning process by learning through experiences, retaining the learned knowledge, and using this knowledge on present tasks. Inspired by the learning of a human brain, we propose a continual learning methodology for point cloud registration which is a key component for 3D reconstruction and Augmented Reality (AR) localization. A major challenge that is faced in this context is the accurate correspondence of 3D points. Current learning-based registration techniques perform substantially, but these methods do not take into account the relations among point clouds obtained during sequential scans and perform registrations among point cloud pairs independently. Our proposed methodology increases the expressiveness of point clouds under consideration by using a novel continual graph network architecture with an attention mechanism that utilizes the learning from the association of the previous point cloud pairs to associate and register points in the current pair. The methodology implemented in PYTHON is evaluated on various challenging benchmark datasets including the Oxford, the 3DMatch, the KITTI, and the ETH dataset. The methodology is thoroughly evaluated under various scenarios including correspondence performance, registration performance, and generalization abilities. The results show a remarkable performance improvement as compared to the state-of-the-art methods. More critically, the proposed method offers exemplary globalization performance across unseen datasets and new scenarios. • Continual graph attention network for point cloud registration. • Brain inspired learning with knowledge consolidation, retention and reusability. • Generalized model capable of estimating registration on unseen data. • Performance analysis for point cloud registration in indoor and outdoor environment. • Significant performance increase as compared to other neural network methods. [ABSTRACT FROM AUTHOR]