Johan A. du Preez

Application-independent evaluation of speaker detection

We propose and motivate an alternative to the traditional error-based or cost-based evaluation metrics for the goodness of speaker detection performance. The metric that we propose is an information-theoretic one, which measures the effective amount of information that the speaker detector delivers to the user. We show that this metric is appropriate for the evaluation of what we call application-independent detectors, which output soft decisions in the form of log-likelihood-ratios, rather than hard decisions. The proposed metric is constructed via analysis and generalization of cost-based evaluation metrics. This construction forms an interpretation of this metric as an expected cost, or as a total error-rate, over a range of different application-types. We further show how the metric can be decomposed into a discrimination and a calibration component. We conclude with an experimental demonstration of the proposed technique to evaluate three speaker detection systems submitted to the NIST 2004 Speaker Recognition Evaluation.

An optimal image transform for threshold-based cloud detection using heteroscedastic discriminant analysis

We present a simple image transform that optimally combines four image channels into a greyscale image for threshold-based cloud detection. These image channels, namely blue, green, red and near infrared, are present on many low Earth-orbit resource satellites. Applying a single threshold to a greyscale image is a computationally efficient method suitable for onboard implementation. We used heteroscedastic discriminant analysis (HDA), which is a generalization of the popular dimension-reducing linear discriminant analysis, to transform the image. Comparative tests between HDA, existing transforms from the remote-sensing literature (the haze optimized and D transforms), as well as the single red and blue image channels were conducted. Although thin clouds remain challenging for global threshold-based techniques, the HDA transform consistently gave the best average segmentation errors across the test dataset. This dataset consisted of 32 1 megapixel Quickbird and Landsat images. HDA has not previously been applied to remote-sensing data.

Left ventricular segmentation from MRI datasets with edge modelling conditional random fields.

BackgroundThis paper considers automatic segmentation of the left cardiac ventricle in short axis magnetic resonance images. Various aspects, such as the presence of papillary muscles near the endocardium border, makes simple threshold based segmentation difficult.MethodsThe endo- and epicardium are modelled as two series of radii which are inter-related using features describing shape and motion. Image features are derived from edge information from human annotated images. The features are combined within a discriminatively trained Conditional Random Field (CRF). Loopy belief propagation is used to infer segmentations when an unsegmented video sequence is given. Powell’s method is applied to find CRF parameters by minimizing the difference between ground truth annotations and the inferred contours. We also describe how the endocardium centre points are calculated from a single human-provided centre point in the first frame, through minimization of frame alignment error.ResultsWe present and analyse the results of segmentation. The algorithm exhibits robustness against inclusion of the papillary muscles by integrating shape and motion information. Possible future improvements are identified.ConclusionsThe presented model integrates shape and motion information to segment the inner and outer contours in the presence of papillary muscles. On the Sunnybrook dataset we find an average Dice metric of 0.91±0.02 and 0.93±0.02 for the inner and outer segmentations, respectively. Particularly problematic are patients with hypertrophy where the blood pool disappears from view at end-systole.

Regularized Gaussian belief propagation

Belief propagation (BP) has been applied in a variety of inference problems as an approximation tool. BP does not necessarily converge in loopy graphs, and even if it does, is not guaranteed to provide exact inference. Even so, BP is useful in many applications due to its computational tractability. In this article, we investigate a regularized BP scheme by focusing on loopy Markov graphs (MGs) induced by a multivariate Gaussian distribution in canonical form. There is a rich literature surrounding BP on Gaussian MGs (labelled Gaussian belief propagation or GaBP), and this is known to experience the same problems as general BP on graphs. GaBP is known to provide the correct marginal means if it converges (this is not guaranteed), but it does not provide the exact marginal precisions. We show that our adjusted BP will always converge, with sufficient tuning, while maintaining the exact marginal means. As a further contribution we show, in an empirical study, that our GaBP variant can accelerate GaBP and compares well with other GaBP-type competitors in terms of convergence speed and accuracy of approximate marginal precisions. These improvements suggest that the principle of regularized BP should be investigated in other inference problems. The selection of the degree of regularization is addressed through the use of two heuristics. A by-product of GaBP is that it can be used to solve linear systems of equations; the same is true for our variant and we make an empirical comparison with the conjugate gradient method.

Graph Coloring: Comparing Cluster Graphs to Factor Graphs

We present a means of formulating and solving graph coloring problems with probabilistic graphical models. In contrast to the prevalent literature that uses factor graphs for this purpose,we instead approach it from a cluster graph perspective. Since there seems to be a lack of algorithms to automatically construct valid cluster graphs,we provide such an algorithm (termed LTRIP). Our experiments indicate a significant advantage for preferring cluster graphs over factor graphs,both in terms of accuracy as well as computational efficiency.

SAWACMMM'17: The 1st Workshop on Multi Media Applications within the South African Context

The South African research community has strong individual interests in pattern recognition and machine learning, but to date has had limited interactions with the worldwide multimedia research community. In an attempt to redress this, this workshop aims to introduce a selection of South African researchers to the multimedia community, and expose the multimedia community to a range of multimedia-related work, primarily from South Africa. The theme to be presented will be broader than strict multimedia, but as far as the methodologies applied are concerned, investigations explored are relevant to real time practical problems encountered in multimedia. Many of the problems to be discussed are also particular to Southern Africa. The applications proposed for discussion vary from biological research to communications protocol optimisation and computer vision. More generally, the workshop will focus on probabalistic graphical models, a powerful modelling technique that we believe will be of great interest to the multimedia research community.