The automatic identification of entities like objects, people or their actions in visual data, such as images or video, has significantly improved, and is now being deployed in access control, social media, online retail, autonomous vehicles, and several other applications. This visual recognition capability leverages supervised learning techniques, which require large amounts of labeled training data from the target distribution representative of the particular task at hand. However, collecting such training data might be expensive, require too much time, or even be impossible. In this work, we introduce several novel approaches aiming at compensating for the lack of target training data. Rather than leveraging prior knowledge for building task-specific models, typically easier to train, we focus on developing general visual recognition techniques, where the notion of prior knowledge is better identified by additional information, available during training. Depending on the nature of such information, the learning problem may turn into domain adaptation (DA), domain generalization (DG), leaning using privileged information (LUPI), or domain adaptation with privileged information (DAPI).;When some target data samples are available and additional information in the form of labeled data from a different source is also available, the learning problem becomes domain adaptation. Unlike previous DA work, we introduce two novel approaches for the few-shot learning scenario, which require only very few labeled target samples, and even one can be very effective. The first method exploits a Siamese deep neural network architecture for learning an embedding where visual categories from the source and target distributions are semantically aligned and yet maximally separated. The second approach instead, extends adversarial learning to simultaneously maximize the confusion between source and target domains while achieving semantic alignment.;In complete absence of target data, several cheaply available source datasets related to the target distribution can be leveraged as additional information for learning a task. This is the domain generalization setting. We introduce the first deep learning approach to address the DG problem, by extending a Siamese network architecture for learning a representation of visual categories that is invariant with respect to the sources, while imposing semantic alignment and class separation to maximize generalization performance on unseen target domains.;There are situations in which target data for training might come equipped with additional information that can be modeled as an auxiliary view of the data, and that unfortunately is not available during testing. This is the LUPI scenario. We introduce a novel framework based on the information bottleneck that leverages the auxiliary view to improve the performance of visual classifiers. We do so by introducing a formulation that is general, in the sense that can be used with any visual classifier.;Finally, when the available target data is unlabeled, and there is closely related labeled source data, which is also equipped with an auxiliary view as additional information, we pose the question of how to leverage the source data views to train visual classifiers for unseen target data. This is the DAPI scenario. We extend the LUPI framework based on the information bottleneck to learn visual classifiers in DAPI settings and show that privileged information can be leveraged to improve the learning on new domains. Also, the novel DAPI framework is general and can be used with any visual classifier.;Every use of auxiliary information has been validated extensively using publicly available benchmark datasets, and several new state-of-the-art accuracy performance values have been set. Examples of application domains include visual object recognition from RGB images and from depth data, handwritten digit recognition, and gesture recognition from video.
展开▼
