Hamimah Ujir

Measuring Task Performance Using Gaze Regions

We present a novel method for measuring task performance using gaze regions, i.e., scene regions fixated by a subject as he or she performs a familiar manual task. The scene regions are learned as a bag of features representation, using library lookup based on the Histogram of Oriented Gradients feature descriptor [1]. By establishing a set of task-specific exemplar models, i.e., models sourced from Pareto optimal sequences, the approach recognizes the local optima within a set of task-specific unlabeled models by estimating the distance (of each unlabeled model) to the exemplar models. During testing, the method is evaluated against a dataset of egocentric sequences, each containing gaze data, belonging to three manual skill-based activities. The results show perfect classifications accuracy on several proposed schemes.

3D Facial Expression Classification Using 3D Facial Surface Normals

With current advanced 3D scanners technology, direct anthropometric measurements are easily obtainable and it offers 3D geometrical data suitable for 3D face processing studies. Instead of using the raw 3D facial points, we extracted one of its derivatives, 3D facial surface normals. We constructed a statistical model for variations in facial shape due to changes in six basic expressions using 3D facial surface normals as the feature vectors. In particular, we are interested in how such facial expression variations manifest themselves in terms of changes in the field of 3D facial surface normals. Using our approach, using 3D facial surface normal yields a better performance than 3D facial points and 3D distance measurements in facial expression classification. The attained results suggest surface normals do indeed produce a comparable result particularly for six basic facial expressions with no intensity information.

Spatio-temporal semantic representation of Cardiac MRI in heart attack patients

Semantic Web technologies, applications and tools have made great steps forward in the life science and health care data exchange. However, developing appropriate semantic representations, including designing spatio-temporal ontologies, remains difficult and challenging. In this paper, we describe a framework to engineer a spatio-temporal semantic representation for the Cardiac MRI images using the current existing case studies conducted in Sarawak General Hospital Heart Centre.

Automatic Segmentation of CMRIs for LV Contour Detection

Research on detecting, recognising and interpreting cardiovascular magnetic resonance images (CMRIs) has started since the 1980s. Time consuming and the need of expert evaluation are the key problems in the manual tracing efforts of CMRIs in a routine investigation. CMRIs manual tracing is also dependent on image quality, and there is no one-size-fits-all MRI setting for an optimum image result. In this paper, we present an approach using 2-Standard Division (2-SD) correlation along with the Sum of Absolute Difference technique and Otsu Watershed to automatically detect the left ventricle (LV) wall and blood pool in the effort to automatically assist the assessment of cardiac function. We test the approach using the Sunnybrook Cardiac Data, a standard benchmark dataset. The results shown that the proposed method had improved the automatic detection of the epicardium and endocardium.

A modular approach and voting scheme on 3D face recognition

In this paper, we carried out a modular approach human 3D face recognition across neutral and six basic facial expressions experiments. Initially, a face model is decomposed into several modules before the 3D facial points for each of the modules are extracted. Three sizes of modules are used in our experiments: 2-Module, 6-Module and 10-Module. We apply Support Vector Machines as the classifier to each of the modules. A Majority Voting Scheme (MVS) and Weighted Voting Scheme (WVS) are constructed to infer the emotion underlying a collection of modules. From the analysis, we conclude that 10-Module outperforms 2-Module and 6-Module. In addition, the modules with low amount feature vectors and only contain boundary feature vectors perform worst.

Facial Expression Recognition Using FAPs-Based 3DMMM

A 3D modular morphable model (3DMMM) is introduced to deal with facial expression recognition. The 3D Morphable Model (3DMM) contains 3D shape and 2D texture information of faces extracted using conventional Principal Component Analysis (PCA). In this work, modular PCA approach is used. A face is divided into six modules according to different facial features which are categorized based on Facial Animation Parameters (FAP). Each region will be treated separately in the PCA analysis. Our work is about recognizing the six basic facial expressions, provided that the properties of a facial expression are satisfied. Given a 2D image of a subject with facial expression, a matched 3D model for the image is found by fitting them to our 3D MMM. The fitting is done according to the modules; it will be in order of the importance modules in facial expression recognition (FER). Each module is assigned a weighting factor based on their position in priority list. The modules are combined and we can recognize the facial expression by measuring the similarity (mean square error) between input image and the reconstructed 3D face model.

A Heuristic Coordinate Filtering Scheme to Aid Online Indexing of Spatio-Temporal Trajectories

Traditional online indexing of spatio-temporal trajectories requires indexing coordinates at fixed intervals or each time the moving object changes direction or velocity. We propose a heuristic coordinate filtering scheme to help evaluate each periodically sampled coordinate candidacy for indexing. Successful coordinates are the minimum trajectorys coordinates required as control points during the cubic spline interpolation process; invoked whenever a spatio-temporal historical query is made on the indexed trajectory. Using our heuristic coordinate filtering scheme, the number of indexed coordinates is significantly reduced to 25% - 35% of the original amount required by the traditional indexing methods, yet the indexed trajectorys integrity is well preserved. Both heuristic coordinate filtering and cubic spline interpolation step require minimal time penalty, hence allowing our method for online uses.