Would be cool to know what approaches they take. Stereo and then perspective scaling (smaller thing are farther away) and then probably a blend of spice of ML for recognizing things/knowing expected scale on the fly... idk I'm arm-chair skeptical of vision only but humans do it.