Why is it that in the paper, rgb is input to get dense depth, while in testing code, rgb, dep, gt is input, which becomes depth completion