Incremental learning through unsupervised adaptation in video face recognition

Resumen

In the last decade, deep learning has brought an unprecedented leap forward for computer vision general classification problems. One of the keys to this success is the availability of extensive and wealthy annotated datasets to use as training samples. In some sense, a deep learning network summarises this enormous amount of data into handy vector representations. For this reason, when the differences between training datasets and the data acquired during operation (due to factors such as the acquisition context) are highly marked, end-to-end deep learning methods are susceptible to suffer performance degradation. While the immediate solution to mitigate these problems is to resort to an additional data collection and its correspondent annotation procedure, this solution is far from optimal. The immeasurable possible variations of the visual world can convert the collection and annotation of data into an endless task. Even more when there are specific applications in which this additional action is difficult or simply not possible to perform due to, among other reasons, cost-related problems or privacy issues. This Thesis proposes to tackle all these problems from the adaptation point of view. Thus, the central hypothesis assumes that it is possible to use operational data with almost no supervision to improve the performance we would achieve with general-purpose recognition systems. To do so, and as a proof-of-concept, the field of study of this Thesis is restricted to face recognition, a paradigmatic application in which the context of acquisition can be especially relevant. This work begins by examining the intrinsic differences between some of the face recognition contexts and how they directly affect performance. To do it, we compare different datasets, and their contexts, against each other using some of the most advanced feature representations available to determine the actual need for adaptation. From this point, we move to present the novel method, representing the central contribution of the Thesis: the Dynamic Ensembles of SVM (De-SVM). This method implements the adaptation capabilities by performing unsupervised incremental learning using its own predictions as pseudo-labels for the update decision (the self-training strategy). Experiments are performed under video surveillance conditions, a paradigmatic example of a very specific context in which labelling processes are particularly complicated. The core ideas of De-SVM are tested in different face recognition sub-problems: face verification and, the more complex, general closed- and open-set face recognition. In terms of the achieved results, experiments have shown a promising behaviour in terms of both unsupervised knowledge acquisition and robustness against impostors, surpassing the performances achieved by state-of-the-art non-adaptive methods.