Weakly-labeled data and identity-normalization for facial image analysis

This thesis deals with improving facial recognition and facial expression analysis using weak sources of information. Labeled data is often scarce, but unlabeled data often contains information which is helpful to learning a model. This thesis describes two examples of using this insight.;Using the results from these two different types of search queries as forms of weak labels, a robust method for classification can be developed. This method is based on graphical models, but also encorporates a probabilistic margin. More specifically, using a model inspired by the variational relevance vector machine (RVM), a probabilistic alternative to transductive support vector machines (TSVM) is further developed. In contrast to previous formulations of RVMs, the choice of an Exponential hyperprior is introduced to produce an approximation to the L1 penalty. Experimental results where noisy labels are simulated and separate experiments where noisy labels from image and video search results using names as queries both indicate that weak label information can be successfully leveraged.;Since the model depends heavily on sparse kernel regression methods, these methods are reviewed and discussed in detail. Several different sparse priors algorithms are described in detail. Experiments are shown which illustrate the behavior of each of these sparse priors. Used in conjunction with logistic regression, each sparsity inducing prior is shown to have varying effects in terms of sparsity and model fit. Extending this to other machine learning methods is straight forward since it is grounded firmly in Bayesian probability. An experiment in structured prediction using Conditional Random Fields on a medical image task is shown to illustrate how sparse priors can easily be incorporated in other tasks, and can yield improved results..;Labeled data may also contain weak sources of information that may not necessarily be used to maximum effect. For example, facial image datasets for the tasks of performance driven facial animation, emotion recognition, and facial key-point or landmark prediction often contain alternative labels from the task at hand. In emotion recognition data, for example, emotion labels are often scarce. This may be because these images are extracted from a video, in which only a small segment depicts the emotion label. As a result, many images of the subject in the same setting using the same camera are unused.;The first is a novel method for face-recognition based on leveraging weak or noisily labeled data. Unlabeled data can be acquired in a way which provides additional features. These features, while not being available for the labeled data, may still be useful with some foresight. This thesis discusses combining a labeled facial recognition dataset with face images extracted from videos on YouTube and face images returned from using a search engine. The web search engine and the video search engine can be viewed as very weak alternative classifier which provide weak labels.".;However, this data can be used to improve the ability of learning techniques to generalize to new and unseen individuals by explicitly modeling previously seen variations related to identity and expression. Once identity and expression variation are separated, simpler supervised approaches can work quite well to generalize to unseen subjects. More specifically, in this thesis, probabilistic modeling of these sources of variation is used to identity-normalize" various facial image representations. A variety of experiments are described in which performance on emotion recognition, markerless performance-driven facial animation and facial key-point tracking is consistently improved. This includes an algorithm which shows how this kind of normalization can be used for facial key-point localization.;In many cases in facial images, sources of information may be available that can be used to improve tasks. This includes weak labels which are provided during data gathering, such as the search query used to acquire data, as well as identity information in the case of many experimental image databases. This thesis argues in main that this information should be used and describes methods for doing so using the tools of probability.