Classification of Clothing Attributes Across Domains

Abstract: Classifying clothing attributes in surveillance images can be useful in the forensic field, making it easier to, for example, find suspects based on eyewitness accounts. Deep Neural Networks are often used successfully in image classification, but require a large amount of annotated data. Since labeling data can be time consuming or difficult, and it is easier to get hold of labeled fashion images, this thesis investigates how the domain shift from a fashion domain to a surveillance domain, with little or no annotated data, affects a classifier. In the experiments, two deep networks of different depth are used as a base and trained on only fashion images as well as both labeled and unlabeled surveillance images, with and without domain adaptation regularizers. The surveillance dataset is new and consists of images that were collected from different surveillance cameras and annotated during this thesis work. The results show that there is a degradation in performance for a classifier trained on the fashion domain when tested on the surveillance domain, compared to when tested on the fashion domain. The results also show that if no labeled data in the surveillance domain is used for these experiments, it is more effective to use the deeper network and train it on only fashion data, rather than to use the more complicated unsupervised domain adaptation method.