Skip to content Skip to navigation
Pre-Defense
6/5/2015 12:00 pm
CBIM Multipurpose Room ( Room 22 )

Hybrid Discriminative-Generative Methods and Applications for Human Pose Reconstruction from Monocular Imagery

Mark Dilsizian, Rutgers University

Defense Committee: Dimitris Metaxas (Chair), Ahmed Elgammal and Casimir Kulikowski

Abstract

Estimating 3D Human pose from monocular images is an important and challenging problem in computer vision with numerous applications including human-computer interaction, human activity recognition, security, and health-care. Existing state-of-the-art methods have been largely data-driven. These learning-based models are inherently limited because they cannot leverage anthropomorphic, kinematic, and other physics-based constraints that help to estimate human pose in 3D space. However, fitting a global and part-based generative model to the entire search space can be computationally prohibitive. We combine discriminative approaches with a generative part-based model for recovering an articulated human pose that gives a globally optimal skeleton over image features. The method guarantees a plausible human pose while also resolving local ambiguities among body parts. Qualitative evaluation of the proposed methods on human pose datasets show improvement in reconstruction accuracy compared to current state-of-the-art methods.