A survey on Monocular 3D Human Pose Estimation

PhD Qualifying Examination


Title: "A survey on Monocular 3D Human Pose Estimation"

by

Mr. Shichao LI


Abstract:

Monocular 3D pose estimation for articulated objects is a typical and 
fundamental inverse problem in computer vision, which has received long lasting 
attention and enabled numerous applications such as action recognition, 
surveillance and human-computer interaction. This survey focuses on estimating 
3D pose for humans from a single camera view, and covers the motivation, a 
formal problem formulation, mainstream and state-of-the-art approaches as well 
as open problems and future research directions.

The survey first presents the significance for studying this subject, after 
which a probabilistic Bayesian framework is employed to formulate the problem 
as inferring 3D pose states from image observations/evidences. Afterwards, a 
categorization of the inference methods into generative approaches and 
discriminative approaches is introduced. To assist the analysis of different 
methods, a discussion on various 3D human pose representations (e.g. 
coordinate, joint-angle and  graphical) and how they encode prior knowledge is 
given. Based on the previous discussion, a review of the mainstream and 
state-of-the-art methods is conducted according to our taxonomy, i.e. how they 
conduct inference and which data representation they use. The review also 
critically identifies their strengths and limitations.

Finally, some open problems, potential solutions and appealing directions are 
pointed out. For example, discussion will be directed to the difficulty of 
obtaining rich 3D pose annotations and how recent deep learning based 
discriminative methods can easily exploit dataset bias, ending up with poor 
generalization to new environments. Our recent exploration on mitigating this 
problem will also be presented.


Date:			Friday, 13 September 2019

Time:                  	10:30am - 12:30pm

Venue:                  Room 3598
                         Lifts 27/28

Committee Members:	Prof. Tim Cheng (Supervisor)
 			Prof. Chi-Keung Tang (Chairperson)
 			Dr. Qifeng Chen
 			Dr. Yu-Wing Tai


**** ALL are Welcome ****