An investigation into locating generalised articulated pose models using self-supervised deep learning

Abstract

This thesis investigates how self-supervised learning can be used to locate generalised three-dimensional articulated poses from images. We split this difficult problem into its constituent components, self-supervised keypoint detection to find two-dimensional keypoints from images, and self-supervised pose lifting to infer the depth of those points. We frame this problem as a representation learning problem, where keypoints are a spatially constrained representation, and also consider the semantic properties of keypoints when applied to different use cases. We consider how priors are used to resolve an ill posed problem such as this, before devising a new prior which leverages the rigidity of limbs found in most articulated objects to both locate better keypoints and to improve the lifting of keypoints. We conclude by describing the wider applicability outside of this specifc approach, and suggest future work that logically follows on from this thesis.

More information

Identifiers

ORCID for Jonathon Hare: orcid.org/0000-0003-2921-4283

Catalogue record

Export record