Back to search

University of British Columbia

2D pose regression using implicit Gaussian mixtures

Abstract

dc:description

The task of inferring human poses in 2D image data is known as 2D human pose estimation. Modern approaches to this problem use neural networks to generate keypoint predictions: estimates for the locations of keypoints, such as elbows and knees, associated with the human body. Top-down heatmap-based methods are one such approach, in which predictions are generated for each person in an image independently. This is often done by generating a set of heatmaps which can then be used to predict keypoint locations. Although these methods perform well in practice, they have several disadvantages: the heatmaps they generate are in general not probabilistic and cannot be used to attribute confidence to keypoint predictions, nor correlation between uncertain joint positions. These methods also cannot generally be used in end-to-end regression from input to predictions, as the argmax function is used to determine keypoint locations from heatmaps is non-differentiable. In this thesis, we investigate ways to overcome these drawbacks, with the aim of providing a top-down 2D human pose estimation method which provides predictions which are multi-modal, probabilistic, and differentiable. We accomplish this by creating a model which can generate multiple Gaussian predictions via sampling, leveraging the effectiveness of heatmap-based approaches by generating these Gaussian predictions from heatmaps. Using a sampling procedure ensures predictions are not limited to a small number of preset modes, as is the case with Gaussian mixture models. Implicit Maximum Likelihood Estimation is also used during training, providing a less costly training procedure than performing maximum likelihood estimation on Gaussian mixture distributions. We show that this model is able to generate pose predictions with comparable accuracy to existing methods, while providing rich statistics on the certainty of its predictions over full poses, unlike traditional method which can only produce predictions for each keypoint independently. It is also able to provide multi-modal predictions, capturing information about ambiguous image data that standard heatmap approaches cannot.

Degree

thesis:*
Name thesis:degree_name
Master of Science - MSc
Level thesis:degree_level
master's
Discipline thesis:degree_discipline
Computer Science
Grantor dc:publisher
University of British Columbia
Year dc:date
2026

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Evans, Andrew
Contributors dc:contributor
  • Sigal, Leonid
  • Rhodin, Helge

Rights

dc:rights
Statement dc:rights
  • Attribution-NonCommercial-NoDerivatives 4.0 International
Language dc:language
eng

Identifiers

dc:identifier.*
Handle dc:identifier
http://hdl.handle.net/2429/94153

Chain of custody

source
Harvested from
University of British Columbia
Base URL
circle.library.ubc.ca/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
related terms
citation

Evans, Andrew. 2D pose regression using implicit Gaussian mixtures. master's thesis, University of British Columbia, 2026. http://hdl.handle.net/2429/94153