Geometric deep learning for 3D shape analysis - Part 1

Michael Bronstein
Imperial College London
USI Lugano

The past decade in computer vision research has witnessed the emergence of deep learning, allowing to automatically learn powerful feature representations from large collections of examples. Deep neural networks achieved a breakthrough in performance in a wide range of applications such as image classification, segmentation, detection and annotation. When attempting to apply deep learning to 3D geometric data, one has to face fundamental differences between images and geometric objects. One of the key differences is that in the geometry processing and computer graphics communities, shapes are modeled as manifolds, and one requires to generalize deep neural networks using intrinsic constructions. Intrinsic deep neural networks have recently been used to learn invariant shape features and correspondence, allowing to achieve state-of-the-art performance in several shape analysis tasks, while at the same time allowing for different shape representations, e.g. meshes, point clouds, or graphs.

The purpose of this tutorial is to overview the foundations and the current state of the art in learning techniques for 3D shape analysis and vision. We will overview deep learning techniques for tasks of shape classification, object recognition, retrieval and correspondence. The tutorial will present in a new light the problems of shape analysis, emphasizing the analogies and differences with the classical 2D setting, and showing how to adapt popular learning schemes in order to deal with deformable objects.

Back to Geometry and Learning from Data Tutorials