Research Project Title:
Selection Policies for Stream-Based Active Learning for DNN Monocular Depth Estimation
abstract:Depth perception is a critical task for autonomous robotics, but dedicated depth sensors are heavy and use lots of processing power. Our research seeks to use Deep Neural Networks to estimate depth from RGB images. Beyond purely estimating depth, we hope to train these networks to also predict uncertainty associated with these predictions. Using these uncertainty estimates we plan to investigate methods to improve prediction performance on-the-fly using active learning techniques, while remaining power efficient.
Participating in SuperUROP will enable me to take a deep dive into a specific area of the robotics field. I have previously done work in the motion planning and decision making spaces, and am excited to learn about perception. My dream is to work on robotic applications in space exploration, where weight and power consumption are critical performance criteria, so I am especially excited to be working in the Low-Energy Autonomous Navigation group!