A striking feature of human vision is our ability to learn a model of a novel object from just a single image and then robustly detect and localize instances of that object in 3D scenes. In my SuperUROP project, I aim to develop a vision system capable of this type of efficient model learning and fast object detection and localization. My approach is based on the vision-as-inverse-graphics paradigm and will use object models capturing coarse 3D shape and appearance, which can be learned from limited data, combined with hybrids of enumerative search-based inference and fast template-matching-based detectors. My goal is to present a few-shot scene understanding system and to study this system’s speed, accuracy, robustness, costs (data, compute) as compared to deep learning approaches.

I enjoyed my project on Bayesian inverse graphics from the past year, and I am confident that the collaborations and mentorship from my UROP supervisors and PIs have meaningfully impacted my MIT experience. By continuing my involvement with the project as a SuperUROP student, I hope to gain broader and deeper understanding of the research field, and to make a unique contribution at the intersection of computer vision and computational cognitive science.