What is the issue with perspective transformation that makes it so that the encoding using a matrix and vector doesn't work?