|Description:||Image retrieval based on CNNs provides impressive performance on daytime natural images. Nevertheless, the performance deteriorates quickly when night images are used to search in a database that contains day images only. This issue is known in machine learning as the task of domain adaptation. In this topic, domain adaptation in image retrieval is addressed by changing the visual domain of the images, translating day images into night images, before processing them by the image retrieval pipeline. A conditional generative adversarial network (GAN) is to be trained in a weakly supervised setup on images from the two domains in order to provide such a translation. The essence of the topic is in the adjustment of the GAN training to profit the downstream task of image retrieval.
To complete the topic, the student is expected to perform research - to study the problem and explore possible solutions. The topic is of interest to the computer vision community and is aimed towards a publication at a computer vision conference.