Annotated reconstruction of 3D spaces using drones

As the fields of robotics and drone technologies are continually advancing, the challenge of teaching these agents to learn and maneuver in the real world becomes increasingly important. A critical component of this is the ability for a robot to map and understand its surrounding unknown environment, both in terms of physical structure and object classification. In this project we tackle the challenge of mapping a 3D space with annotations using only 2D images acquired from a Parrot Drone. In order to make such a system operate efficiently in close to real time, we address a number challenges including (1) creating a optimized version of Faster RCNN that can operate on drone hardware while still being accurate, (2) developing a method to reconstruct 3D spaces from 2D images annotated with bounding boxes, and (3) using generated 3D annotations to complete drone motion planning for unknown space exploration.