Research

Analysis of fibrous structures in 3-D images: Orientation estimation, bifurcation detection and tractography

The objective is to create 1) directional maps of fibrous structures, 2) robustly detect local complexities such as crossings or branching points, and 3) devise novel deterministic and stochastic techniques for fiber tracking. Specifically, we investigate structures such as microtubules, cardiac myofibers,  free-running cardiac Purkinje system, coronary arteries, etc. Our methods involve nonlinear filtering techniques, robust branch detectors, and analysis of orientation distribution functions (ODFs).

Unsupervised techniques for manifold clustering

We investigate an alternative mean shift formulation, which performs the iterative optimization ``on'' the manifold of interest and intrinsically locates the modes via consecutive evaluations of a mapping. In particular, these evaluations constitute a modified gradient ascent scheme on Stiefel and Grassmann manifolds. The performance evaluated by conducting experiments on object categorization and segmentation of multiple motions.

Graph theoretic image & motion segmentation

Graph theoretic algorithms have attracted widespread interest from the computer vision community since they provide powerful ways to solve both image and motion segmentation problems. In this project, we investigate multi-class, semi-automatic image/motion segmentation algorithms based on random walks on 2-D weighted graphs. We further analyze the transient behavior of the network, which allow to incorporate different features such as texture, shape, etc. in addition to intensity/optical flow.

Fiber clustering & segmentation in DTI

In Diffusion Tensor Imaging (DTI), the data lives on a 6-dimensional space with nontrivial geometry, which encodes the orientation of the anisotropy. In segmentation problem we tey to group fibers into tracts. The first step is to define (dis)similarity metrics on the space of tensors. We are investigating graph theoretical methods for fiber clustering and segmentation.

Dynamical system theoretic lip articulation analysis

We present a system for synthesizing lip movements and recognizing speakers/phrases from visual lip sequences. The temporal evolution of the lip features is modeled with linear dynamical systems (LDS), whose parameters are learned using system identification techniques. By carefully exploiting physical constraints of lip movement both in the learning and synthesis stages, realistic synthesis of novel sequences based on the learned LDS is achieved. Recognition is performed using classification methods, such as nearest neighbors and support vector machines, combined with metrics for dynamical systems, such as subspace angles and Binet-Cauchy trace kernels.

Discriminative lip motion features for multimodal speaker identification & speech-reading (at MVGL-KOC)

We present a new multimodal speaker/speech recognition system that integrates audio, lip texture and lip motion modalities. We propose using the explicit lip motion information, instead of or in addition to lip texture and/or geometry information within a unified feature selection and discrimination analysis framework. A novel two-stage, spatial and temporal discrimination analysis is also introduced to select the best lip motion features. It solves the dimension reduction problem optimally, taking into account the intra-class and inter-class distribution of individual single-frame lip feature vectors as well as the temporal discrimination information. We then investigate the benefits of inclusion of the best, i.e. the most discriminative lip motion modality for multimodal recognition. The fusion of audio, lip texture and lip motion modalities is performed by the so-called Reliability Weighted Summation (RWS) decision rule.