Beyond the Complete-Body Prior: Towards Robust Human Perception under Severe Occlusions and Atypical Topologies

Beyond the Complete-Body Prior: Towards Robust Human Perception under Severe Occlusions and Atypical Topologies

Speaker: Qingzheng Xu

Abstract:

Human understanding in computer vision goes beyond simply detecting people; it involves interpreting body kinematics, geometric structures, and semantic motion. However, current perception models often rely on standard priors. They typically assume that human targets are fully observable and conform to fixed skeletal templates. When deployed in uncontrolled real-world environments, these standard assumptions can break down. Severe interactive occlusions in dense multi-person scenarios disrupt joint tracking. Simultaneously, atypical physical topologies cause standard 3D mesh regressors to hallucinate intact limbs, and lead language models to misinterpret movement.

 

His seminar explores methods to address these observational and structural bottlenecks. It will discuss the motivation and initial progress in building benchmarks, adaptive 3D mesh recovery frameworks, and topology-agnostic motion interpretation systems. Together, these works aim to advance reliable human perception across geometric, evaluative, and semantic layers for complex real-world applications.

 

Bio:

Qingzheng Xu is a Ph.D. student at the School of EECS, University of Queensland, under the supervision of A/Prof Xin Yu and A/Prof Sen Wang. His research focuses on human perception under severe occlusions and atypical topologies.