Current image recognition algorithms developed for self-driving cars rely primarily on deep learning. Such models must be trained on large datasets to perform well, but datasets of well-annotated images can be limited in size and expensive to acquire. The goal of my project is to use 3D graphics to create large, annotated, synthetic datasets of city scenes to train image-recognition algorithms for self-driving cars. Specifically, I will be working on diversifying rules for generating buildings, roads, and people, and making the rendered city scenes more realistic. The flexibility of a synthetic dataset also allows researchers to study other aspects of deep earning, such as a model’ s capability to generalize, and the effects of image label quality on a model’ s performance.

By participating in SuperUROP, I hope to gain more substantial research experience, learn more about computer graphics, and familiarize myself with current technology and software. I have previously taken computer graphics and machine learning courses, and I am excited to apply what I’ ve learned to a more long-term project.