For example, if a door has mirrors on it, then the model's prediction is as if the door's body is hollow with depth volume. This is similar to the red-eyes problem, which we used to see back in the day with gen1 digi cameras. I am confident a fix should be imminent.