We have begun development of a Machine Learning algorithm for 3D object detection that provides information about an object’s pose, rotation and bounding boxes in AR space, inspired by Google’s Objectron.
This algorithm aims to provide automatic 3D detection of relevant objects, and mapping of those objects to a GIS in real time.
For purposes of training the network, we are developing a 3D Object Annotation tool using Unity, that leverages data captured by an AR device (Point Cloud, RGB image) to train the algorithm to not only detect objects on a video, but also determine its real world position and orientation, all without any user input.
Preliminary annotation tests yielded great results, producing an output suitable for training – the following phase which will begin as soon as enough data is gathered. Big updates are coming, and we’ll keep you informed! 🙂