I'm going to guess this is because the image to depth data, while good, is not perfectly accurate and therefore cannot be a shared ground truth between multiple images. At that point what you want is a more traditional structure from motion workflow, which already exists and does a decent job.