Unpaired learning from irregular geometric modalities

MIT PI: Justin Solomon, Department of Electrical Engineering & Computer Science
SkT PI: Evgeny Burnaev, Center for Computational and Data-Intensive Science and Engineering

Learning from geometric data is a critical task for key applications, including computer-aided design (CAD), additive manufacturing, autonomous driving, molecular biology, and processing of satellite data. Geometric data, however, distinguishes itself from other modalities in computer vision and machine learning---especially images and collections of data points---thanks to its irregularity and sparse structure. Moreover, three-dimensional shapes are rarely accompanied by rich annotations more common for images. Motivated by these considerations, we aim to advance state-of-the-art models and algorithms for learning from geometric modalities like the ones above. Our focus is specifically on algorithms that require little to no supervision, reflecting the realities of readily available geometric data.
This project builds on long-standing interactions between the two PIs’ research groups and reflects core areas of research for both teams. In particular, it builds on a multitude of 3D datasets collected by the Skoltech team as well as learning algorithms for point clouds, CAD data, geographic systems, and vector graphics developed by both teams. The PIs have visited each others’ teams at Skoltech and MIT multiple times for extended research interactions in the past, and this project will deepen and expand the connections between their teams through both joint research and in-person interaction.

Back to the list >>