Lei Zhang







An Extended Chain Graph and its Applications in Computer Vision

Both undirected graphical models and directed acyclic graphical models have some limitation on their modeling power. While Chain Graph models are powerful enough, existing Chain Graph models used in real applications have simplified structures, therefore not fully leveraging the power of Chain Graph. This fact is due to several reasons such as  the difficulty of parameter learning for a complex Chain Graph with very general topology and the lack of efficient probabilistic inference method for such Chain Graph. To this end we developed an extended Chain Graph model and presented a suite of methods for constructing, learning, and inference this model. We successfully applied our theory to two important computer vision applications: multi-subject human activity recognition and 2D image segmentation. More information is available here.


A Bayesian Network Model for Automatic and Interactive Image Segmentation

We proposed a hierarchical Bayesian Network (BN) model for both automatic and interactive image segmentation. Our BN segmentation model captures the statistical relationships among multiple image entities and encodes several constraints for effective image segmentation. In addition, we developed a mutual information based actively interactive segmentation approach that were more effective and efficient than the passively interactive segmentation. More information is available here.


2D and 3D Upper Body Tracking with the Same Framework

Many existing works use different models for 2D and 3D body tracking. In addition, they enforce limited types of constraints (typically the kinematic constraint) between the body parts. In contrast, we developed a single framework that can be used for both 2D and 3D body tracking. In addition, we encoded multiple types of constraints (e.g., anatomical constraints) into the model for effective tracking. More information is available here.


Active Image Labeling and its Applications in Action Unit Labeling

Active labeling aims at minimizing the human labeler's labor and maximizing the efficiency of manual labeling. We use a mutual information based strategy for active image labeling and apply it to efficient labeling of Action Units (AUs) for facial expression recognition study. More information is available here. 


Level Set-based Image Segmentation with Global Shape Prior

For image segmentation of objects with certain types of shapes, the global shape prior information can be leveraged to improve the effectiveness of image segmentation. We propose a global shape based on the contour difference and integrate it into a level-set based image segmentation framework. We have applied this segmentation approach to accurately segment membrane images for biological study. More information is available here.


Video Segmentation using a Spatio-temporal Conditional Random Field

For segmentation in video sequences, the spatial and temporal information among the neighborhood of a pixel should be taken into account. We developed a Spatio-temporal Conditional Random Field (STCRF) model that can effectively capture the spatio-temporal relationships for pixelwise image labeling. More information is available here.  


Side Notes:

I drew an ontology graph that summarizes many existing types of segmentation approaches.  More information is available here.  In addition, in industry I did research on multi-modal hybrid sensor fusion, terrain surface material classification for aerial imagery, physiological health monitoring for first responders, affective computing, visual knowledge discovery as well.