Detecting Waterborne Debris with Sim2Real and Randomization

Marine debris pollution is one of the most ubiquitous and pressing environmental issues affecting our oceans today. Clean up efforts such as the Great Pacific Garbage Patch project have been implemented across the planet to combat this problem. However, resources to accomplish this goal are limited, and the afflicted area is vast. To this end, unmanned vehicles that are capable of automatically detecting and removing small-sized debris would be a great complementary approach to existing large-scale garbage collectors. Due to the complexity of fully functioning unmanned vehicles for both detecting and removing debris, in this project, we focus on the detection task as a first step. From the perspective of machine learning, there is an unfortunate lack of sufficient labeled data for training a specialized detector, e.g., a classifier that can distinguish debris from other objects like wild animals. Moreover, pre-trained detectors on other domains would be ineffective while creating such datasets manually would be very costly. Due to the recent progress of training deep models with synthetic data and domain randomization, we propose to train a debris detector based on a mixture of real and synthetic images.

Speaker bio: Kris is a postdoc at Mila working with Yoshua Bengio on problems related to Humanitarian AI. He is generally interested in ways to broaden the scope of problems studied by the machine learning community and am curious about the ways to bridge statistical and computational thinking.