Open Access

Vedhas Pandit

• Today, the relevance of Affective Computing, i.e., of making computers recognise and simulate human emotions, cannot be overstated. All technology giants (from manufacturers of laptops to mobile phones to smart speakers) are in a fierce competition to make their devices understand not only what is being said, but also how it is being said to recognise user’s emotions. The goals have evolved from predicting the basic emotions (e.g., happy, sad) to now the more nuanced affective states (e.g., relaxed, bored) real-time. The databases used in such research too have evolved, from earlier featuring the acted behaviours to now spontaneous behaviours. There is a more powerful shift lately, called in-the-wild affect recognition, i.e., taking the research out of the laboratory, into the uncontrolled real-world. This thesis discusses, for the very first time, affect recognition for two unique in-the-wild audiovisual databases, GRAS2 and SEWA. The GRAS2 is the only database till date with time-Today, the relevance of Affective Computing, i.e., of making computers recognise and simulate human emotions, cannot be overstated. All technology giants (from manufacturers of laptops to mobile phones to smart speakers) are in a fierce competition to make their devices understand not only what is being said, but also how it is being said to recognise user’s emotions. The goals have evolved from predicting the basic emotions (e.g., happy, sad) to now the more nuanced affective states (e.g., relaxed, bored) real-time. The databases used in such research too have evolved, from earlier featuring the acted behaviours to now spontaneous behaviours. There is a more powerful shift lately, called in-the-wild affect recognition, i.e., taking the research out of the laboratory, into the uncontrolled real-world. This thesis discusses, for the very first time, affect recognition for two unique in-the-wild audiovisual databases, GRAS2 and SEWA. The GRAS2 is the only database till date with time- and value-continuous affect annotations for Labov effect-free affective behaviours, i.e., without the participant’s awareness of being recorded (which otherwise is known to affect the naturalness of one’s affective behaviour). The SEWA features participants from six different cultural backgrounds, conversing using a video-calling platform. Thus, SEWA features in-the-wild recordings further corrupted by unpredictable artifacts, such as the network-induced delays, frame-freezing and echoes. The two databases present a unique opportunity to study time- and value-continuous affect estimation that is truly in-the-wild. A novel ‘Evaluator Weighted Estimation’ formulation is proposed to generate a gold standard sequence from several annotations. An illustration is presented demonstrating that the moving bag-of-words (BoW) representation better preserves the temporal context of the features, yet remaining more robust against the outliers compared to other statistical summaries, e.g., moving average. A novel, data-independent randomised codebook is proposed for the BoW representation; especially useful for cross-corpus model generalisation testing when the feature-spaces of the databases differ drastically. Various deep learning models and support vector regressors are used to predict affect dimensions time- and value-continuously. Better generalisability of the models trained on GRAS2 , despite the smaller training size, makes a strong case for the collection and use of Labov effect-free data. A further foundational contribution is the discovery of the missing many-to-many mapping between the mean square error (MSE) and the concordance correlation coefficient (CCC), i.e., between two of the most popular utility functions till date. The newly invented cost function |MSE_{XY}/σ_{XY}| has been evaluated in the experiments aimed at demystifying the inner workings of a well-performing, simple, low-cost neural network effectively utilising the BoW text features. Also proposed herein is the shallowest-possible convolutional neural network (CNN) that uses the facial action unit (FAU) features. The CNN exploits sequential context, but unlike RNNs, also inherently allows data- and process-parallelism. Interestingly, for the most part, these white-box AI models have shown to utilise the provided features consistent with the human perception of emotion expression.…