Image classification over unknown and anomalous domains

Abstract

A longstanding goal in computer vision research is to develop methods that are simultaneously applicable to a broad range of prediction problems. In contrast to this, models often perform best when they are specialized to some task or data type. This thesis investigates the challenges of learning models that generalize well over multiple unknown or anomalous modes and domains in data, and presents new solutions for learning robustly in this setting.

Initial investigations focus on normalization for distributions that contain multiple sources (e.g. images in different styles like cartoons or photos). Experiments demonstrate the extent to which existing modules, batch normalization in particular, struggle with such heterogeneous data, and a new solution is proposed that can better handle data from multiple visual modes, using differing sample statistics for each.

While ideas to counter the overspecialization of models have been formulated in sub-disciplines of transfer learning, e.g. multi-domain and multi-task learning, these usually rely on the existence of meta information, such as task or domain labels. Relaxing this assumption gives rise to a new transfer learning setting, called latent domain learning in this thesis, in which training and inference are carried out over data from multiple visual domains, without domain-level annotations. Customized solutions are required for this, as the performance of standard models degrades: a new data augmentation technique that interpolates between latent domains in an unsupervised way is presented, alongside a dedicated module that sparsely accounts for hidden domains in data, without requiring domain labels to do so.

In addition, the thesis studies the problem of classifying previously unseen or anomalous modes in data, a fundamental problem in one-class learning, and anomaly detection in particular. While recent ideas have been focused on developing self-supervised solutions for the one-class setting, in this thesis new methods based on transfer learning are formulated. Extensive experimental evidence demonstrates that a transfer-based perspective benefits new problems that have recently been proposed in anomaly detection literature, in particular challenging semantic detection tasks.