· · · Institutional Repository

Automatic analysis of facial expressions is a complex area of pattern recognition and computer vision with many un- resolved problems, one of which is the distinction between posed and spontaneous expressions of emotions. Previous psychology research indicates that the temporal dynamics in the face are essential for distinguishing between posed and spontaneous smiles. There are six temporal characteristics which are important: morphology, apex overlap, symmetry, total duration, speed of onset and speed of offset. In this work, we propose to distinguish between posed and spon- taneous expressions by using Dynamic Bayesian networks (DBN) to model the temporal dynamics. The DBN provides a suitable framework to represent probabilistic relationships between and within the various types of temporal dynamics. Based on the temporal phases of four different Action Units (onset, apex offset and neutral of facial actions) and the six temporal characteristics from the psychology research, we build several DBN models to distinguish between posed and spontaneous expressions. We present experimental results from 50 videos displaying posed and spontaneous smiles. When the DBNs trained on the temporal characteristics are combined to provide a joint classification, we attain an AUC of 0.97.

Dit stageverslag is ter afsluiting van het bachelorproject voor de opleiding Technische Informatica aan de Technische Universiteit Delft. Gedurende de stage is het programma Streampaper ontworpen en geïmplementeerd. Streampaper is een op zichzelf staande applicatie waarin verschillende externe informatiestromen gecombineerd kunnen worden weergegeven. De gebruiker van Streampaper kan gemakkelijk door pagina's met berichten uit deze bronnen bladeren. Het doel van Streampaper is het vergemakkelijken van het up-to-date blijven van werknemers wat betreft bedrijfsnieuws en eventuele andere relevante nieuwsbronnen.

We present a framework for learning arithmetic expressions from a set of observations. Our intention is to introduce a Bayesian method for what is known as equation discovery. Our method is based on measuring a degree of belief (posterior probability) for a set of hypothesized expressions to find those which best explain the observed data. This measure is used as the basis for choosing one hypothesis over another. In our work we distinguish two tasks in the process of equation discovery, namely: the task of exploring the space of arithmetic expressions and that of evaluating the degree that an expression describes the data. Separating these two, allows us to investigate them independently. For the first task, we use a context-free grammar to construct a large set of expressions which we take as our hypothesis space. The set contains a large number of hypotheses (each an arithmetic expression) that should be tested against the data. We also evaluate complexity of for each expression using the grammar. The complexity is presented to the model in the form of a prior probability. Our main focus here is the second task: the posterior evaluation using a Bayes formulation. The method tests a hypothesized expression against a set of provided samples that have quantitative input features. It calculates a likelihood probability which expresses the degree that a hypothesis describes the data. A final posterior probability is calculated based on the prior and the likelihood, that is the measure of qualification for each expression.

Robust automatic detection of surface and air objects in a maritime environment is a problem that is of growing importance to the Royal Netherlands Navy (RNLN). Due to a shift in the field of operation from the open oceans towards the littoral waters, the RNLN is forced to operate in complex environments with cluttered backgrounds and the presence of many small vessels and a wide range of other objects. Traditional radar systems are not optimal in these circumstances due to their minimum detection range, lack of sensitivity to small, non-metallic, objects and poor classification power. Complementation by Electro-Optical (EO) camera systems is therefore desired, which resulted in the start of the development of a detection algorithm based on polynomial background estimation. Automated object detection in the maritime environment is a complex problem however, due to various complicating factors. These factors include the highly dynamic background, camera motion, the variety in possible objects and their appearance, and the diversity in meteorological as well as environmental circumstances. Although the developed detection algorithm is quite well capable of detecting the objects, it also produces an extensive amount of false detections. This study investigates whether these false detections can be eliminated, while maintaining the true detections, by means of classification of the detections as either target or background. To this end, the initial detection algorithm is optimised to detect as much objects as possible in a carefully constructed dataset of eight hundred Visible Light (VL) images. The resulting detections from the optimised algorithm are used accordingly to train and test various basic classifiers, using a set of features found in the literature. The best performing classifier is selected and the performance of this classifier, and the two-stage detection algorithm as a whole, is subsequently further analysed by means of various tests involving the features used, the evaluation procedure and the fusion of detection results. Results show that especially the features as well as the clustering procedure for detected pixels are important parameters with respect to a good performance of the algorithm. This works shows that the linear discriminant classifier is best to use with the problem among the classifiers considered. Moreover, it is demonstrated that including features of histogram equalized boxes in combination with features of the entire image increased the performance the most, that determining the features on a slightly larger area than the originally detected area is beneficial and that fusion of detections after classification can be used to optimise the detector output. Although the developed classification approach is capable of eliminating many false detections and to retain a majority of the true detections, further research is required. Suggested are separate classifiers for the sea- and sky part, inclusion of the time dimension, optimisation of the operating point of the classifier and preprocessing steps.

An important component of question answering systems is question classification. The task of question classification is to predict the entity type of the answer of a natural language question. Question classification is typically done using machine learning techniques. Different lexical, syntactical and semantic features can be extracted from a question. In this work we introduce two new semantic features which improve the accuracy of classification. Furthermore, we developed a weighed approach to optimally combine different features. We also applied Latent Semantic Analysis (LSA) technique to reduce the large feature space of questions to a much smaller and efficient feature space. We adopted two different classifiers: Back-Propagation Neural Networks (BPNN) and Support Vector Machines (SVM). We found that applying LSA on question classification can not only make the question classification more time efficient, but it also improves the classification accuracy by removing the redundant features. Furthermore, we discovered that when the original feature space is compact and efficient, its reduced space performs better than a large feature space with a rich set of features. In addition, we found that in the reduced feature space, BPNN performs better than SVMs which are widely used in question classification. We tested our proposed approaches on the well-known UIUC dataset and succeeded to achieve a new record on the accuracy of classification on this dataset.

Object Tracking has been a very active area in the field of C omputer Vision. Over the years, a variety of approaches have been put forth to solve this problem and though many of them have demonstrate considerable success none of them have been completely successful. With more methods being written each day, the evaluation of such systems becomes a very important task. If an evaluation system exists that is able to point out specific flaws in the stage of development, it can lead to a very robust and improved algorithm. This work attempts to create such an evaluation framework. Given an algorithm that detects people and simultaneously tracks them, we evaluate its output by considering the complexity of the input scene. Some videos used for the evaluation are recorded using the Kinect sensor and a benchmark dataset from the PETS workshop is also used. To analyze the performance of the tracking system,the reasons due to which the algorithm might fail are investigated and quantified over the entire video sequence. A set of features called Scene C omplexity Measures are obtained for each input frame. The variability in the algorithm performance is modeled by these complexity measures using various regression models. From the regression statistics, we show that we can compare the performance of two different algorithms and also quantify the relative influence of the scene complexity measures on a given algorithm.