April 9 - Doctoral Seminar: Prashant Raina

Speaker: Prashant Raina

Supervisors: Drs. S.P. Mudur, T. Popa

Supervisory Committee: Drs. T. Fevens, A. Krzyzak, H. Rivaz

Title: A Parallel Computing Architecture for High-Performance OWL Reasoning

Date: Thursday, April 5, 2018

Time: 10:15 a.m.

Place: EV 3.309

ABSTRACT

While deep learning and convolutional neural networks (CNNs) have shown great success in 2D image processing tasks, there is still significant scope for research in the intersection of deep learning and 3D graphics. Most work in this area has focused on classification tasks using 3D volumetric grids or multiple 2D views of a shape, both of which are Euclidean spaces. Among non-Euclidean surface representations, point clouds are an especially important domain for research, since they are upstream from 3D meshes in the computer vision pipeline. They are also far more difficult to work with due to their lack of structural information. We have developed a method to create a grid-like structure in neighborhoods of point clouds by using Moving Least-Squares (MLS) interpolation to resample the points. This effectively enables CNNs to be applied to local feature learning in point clouds. We then use this method to train a CNN to detect sharp features in point clouds. We show that such results are difficult to achieve with PointNet, the main model behind other contemporary approaches to deep learning in point clouds. To illustrate the utility of the detected sharp features, we have utilized them in a novel feature-aware smoothing method which smooths point clouds while preserving the detected sharp features. The results obtained are competitive with the widely-used Robust Implicit Moving Least-Squares (RIMLS) smoothing method. Future work will use recent developments in geometric deep learning to expand the potential applications for our deep learning method.