3D Multi-bodies: Fitting Sets of Plausible 3D Human Models to Ambiguous Image Data

TL;DR

This paper introduces a neural network approach that generates multiple plausible 3D human reconstructions from ambiguous or occluded images, leveraging a generative model to ensure realistic poses.

Contribution

It proposes a multi-hypothesis neural network with a best-of-M loss for improved 3D human shape and pose estimation under ambiguity and occlusion.

Findings

Outperforms existing methods in ambiguous pose recovery

Effective on heavily occluded data

Utilizes a generative model to constrain plausible human shapes

Abstract

We consider the problem of obtaining dense 3D reconstructions of humans from single and partially occluded views. In such cases, the visual evidence is usually insufficient to identify a 3D reconstruction uniquely, so we aim at recovering several plausible reconstructions compatible with the input data. We suggest that ambiguities can be modelled more effectively by parametrizing the possible body shapes and poses via a suitable 3D model, such as SMPL for humans. We propose to learn a multi-hypothesis neural network regressor using a best-of-M loss, where each of the M hypotheses is constrained to lie on a manifold of plausible human poses by means of a generative model. We show that our method outperforms alternative approaches in ambiguous pose recovery on standard benchmarks for 3D humans, and in heavily occluded versions of these benchmarks.