Comment by polytronic

My understanding is that it generates parallel to the rgb information, a depth-map and segments encoded also in specific spectrum of colors. This removes the burden of heavy image processing (edge-detection, etc) which improves performance. The segments can be trained for classification.