Deep Learning on Geometry Representations

Par

Dmitriy (Dima) Smirnov

MIT

Résumé:

While deep learning has been successfully applied to many tasks in computer graphics and vision, standard learning architectures often operate on shape representations that are dense and regular, like pixel or voxel grids. On the other hand, decades of computer graphics and geometry processing research have resulted in specialized algorithms, tools, and techniques that use representations without such regular structure. In this talk, I will show how revisiting conventional approaches can yield deep learning pipelines and inductive biases that are directly compatible with common geometry representations. In particular, I will discuss works on applying learning to triangle meshes, CAD-style parametric primitives, sprites, and hybrid explicit/implicit shapes with boundaries.

Biographie :

Dmitriy (Dima) Smirnov is a final-year PhD student in the Geometric Data Processing group at MIT, advised by Professor Justin Solomon. His research lies on the intersection of computer graphics, geometry processing, and deep learning. Dima is an NSF Graduate Research Fellow and has done internships at Adobe Research and Pixar Research.