Dominic Spata, M.Sc.

My name is Dominic Spata and I am a machine learning engineer at Aptiv, where I research the use of neural networks for automotive radar perception. I am also currently completing my Ph.D. degree at the Bergische Universität Wuppertal.

I graduated with a master's degree from the Applied Computer Science programme at Ruhr University Bochum. The focus of my studies were machine learning and artifical intelligence, especially as applied to problems from the domains of driving and traffic. Hence, I often worked closely with the Institut für Neuroinformatik, where I also completed both my bachelor's and master's theses.

Currently, I live in Witten, Germany. My interests and hobbies include (but are not limited to) exercising (swimming, biking, and more), reading and writing fantasy, games of all forms (be it sports, video games, board games, or card games), web and graphics design¹, as well as, of course, anything and everything related to programming.

¹This website was created by myself from scratch in pure HTML and CSS.

Title: Action Classification Using a Combination of Hough Voting and Random Forest

Advisors: PD Dr. Rolf Würtz, Dipl.-Inform. André Ibisch

This thesis deals with the classification of human behaviour using a recent method. Given a video of a single human performing some distinct action as well as annotations indicating that human's position, it extracts 3D feature patches based on the dense visual flow of the image material. A random forest regressor is used to estimate the probability of a certain action being localised at a certain position and time in the video. These probabilities are consolidated using a Hough voting scheme to yield the final classification result.

Title: Generation of Natural Traffic Sign Images Using Domain Translation with Cycle-Consistent Generative Adversarial Networks

Advisors: Jun.-Prof. Sebastian Houben, Daniela Horn M.A. M.Sc.

This thesis deals with the generation of novel traffic sign images with the goal of supplementing the German Traffic Sign Benchmark dataset. A custom two-step method for image generation is explored, which first derives prototype images from standardised traffic sign diagrams and then enhances theses prototypes into natural images using an unpaired image-to-image translation system. The translation architecture of choice is a recent invention by the name of cycle-consistent generative adversarial networks (CycleGANs), which combine the powerful generative adversarial framework with CNN-based image-to-image mappings.