Very impressive. Is it possible to combine this with a semantic analyzer in some way? Seems like having a first pass through one and then serving that output on

Very impressive. Is it possible to combine this with a semantic analyzer in some way? Seems like having a first pass through one and then serving that output on top of raw image pixels could reduce the potential color space and maybe blunt training set biases? To wit, I know nothing about these things.