Improving 3D human pose estimation in-the-wild

There has been some suspicion that 3D human pose estimation produces significantly worse results on in-the-wild images than on lab images. Confirming this suspicion is difficult, because it is hard to get 3D ground truth for in-the-wild images without measurement equipment significantly affecting the imagery. This thesis (a) demonstrates the suspicions are correct; (b) shows the effect is, at least in part, due to reconstructions not plausible (that is, "like" human poses); (c) explores simple augmentation can improve performance in situations with occlusion and (d) shows that natural methods to produce reconstructions that are plausible produce measurable improvements for in-the-wild reconstruction. Forcing methods to produce reconstructions that are plausible produces no major improvement on Human3.6M validation data; but this is because error on Human3.6M validation data is a poor predictor of error on in-the-wild data. This thesis shows that a registration error measure applied to reconstructions from multiple view data is a good predictor of ground truth error. Our registration error confirms that various procedures to enforce plausible reconstructions make notable improvements on in-the-wild error consistently across a number of distinct multiple view human action datasets.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2022

Author dc:creator

• Gonzalez, Victor

Contributors dc:contributor

• Forsyth, David A

• human pose estimation
• occlusion

Statement dc:rights

• Copyright 2022 Victor Gonzalez

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/115430

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F115430