Dissertation Research:
Title: The StarGANv2-based generative framework for data augmentation in the domain of eye images
Presenter: Dmytro Katrychuk
Advisor: Oleg Komogortsev
Date/Time: Thursday, December 8th @10:00 a.m.
Location: https://txstate.zoom.us/j/91243653692
Abstract: Data augmentation is a powerful tool for increasing performance and improving robustness of end-to-end deep learning architectures. In this work, we focus on creating a generative framework for the domain of near-eye monocular images. We propose a hybrid solution that fuses two most common approaches from the relevant literature: synthetic rendering and end-to-end learning from large pool of data. This solution enables us to map the real subject identities from pre-collected video-oculography data onto a Blender artificial model, which allows precise control over the rendering scene while severely lacking in diversity and photorealism. After posing this problem as a domain transfer, we research the applicability of a StarGANv2 architecture – a proven solution in that field. We apply the vanilla method and discuss the potential novel improvements specific to the eye-tracking domain. The generated images are assessed in terms of their photorealism, visual and content consistency. The results are benchmarked against the EyeGAN and MUNIT architectures. The proposed method is finally evaluated as a data augmentation tool for increasing the train set of an appearance-based gaze estimation neural network.
Deadline: Jan. 4, 2023, midnight