Structured Generative Models for 3D Vision

Abstract:

Modern techniques for 3D vision tasks typically require large amounts of labelled training data. In this talk, I shall discuss an alternative -- unsupervised learning using structured generative models. Like traditional generative models (e.g. VAEs or denoising diffusion), these structured models learn the distribution of a dataset and can generate new samples. Unlike traditional models, they learn an interpretable latent space and decoder, which explicitly represents positions and 3D shapes of the objects visible in the image. I shall discuss how this allows predicting 3D shape, object segmentations, and depth maps from a video or even a single image, without receiving supervision for any of those tasks.

Speaker:
Dr Paul Henderson is a Lecturer (aka Assistant Professor) in Machine Learning at the University of Glasgow. His research focuses on building machines that understand the visual world with minimal supervision, learning aspects of its structure such as 3D geometry and decomposition into objects. This work draws on techniques from machine learning (particularly deep generative models), computer vision, and computer graphics. Before coming to Glasgow, he completed his PhD with Prof. Vittorio Ferrari at the University of Edinburgh, then worked as a postdoc at ISTA in Vienna, Austria.