We optimized the training, annotation and deploy infra to minimize the time it takes to bring up a custom object detector. The current version only supports detection of object centers (as opposed to full bounding boxes) and tends to do best if the objects don’t vary too much in size.