Suryam Sharma

Graduation Semester and Year

2021

Language

English

Document Type

Thesis

Degree Name

Master of Science in Computer Science

Department

Computer Science and Engineering

First Advisor

Vassilis Athitsos

Abstract

Pose estimation using Deep Neural Networks (DNNs) has shown outstanding performance in recent years, due to the availability of powerful GPUs and larger training datasets. However, there are still many challenges due to the large variability of human body appearances, lighting conditions, complex background, occlusions and postures. Among all these peculiarities, partial occlusions, and overlapping body poses often result in deviated pose predictions. These circumstances can result in wrong and sometimes unrealistic results. The human mind can predict such poses because of the underlying structural awareness of the geometry, of a human body. In this thesis, we discuss an efficient training technique that helps us to correct structurally implausible poses caused due to partial occlusions. We introduce a pose discriminator which helps us to incorporate priors about the human body's structure, into our model. As shown in the experiments, using this pose discriminator results in improved accuracy.

Keywords

Adversarial pose, Human pose estimation, Adversarial PoseNet, Structure aware, Generative adversarial, GAN, Adversarial learning

Disciplines

Computer Sciences | Physical Sciences and Mathematics

Comments

Degree granted by The University of Texas at Arlington

Download

Share

COinS