I bet they are using an implementation of mask rcnn/fast rcnn object detection and bounding box to achieve their objectives.