A "decent" visual recognition API at this point isn't very good or useful. The other comment in this thread is correct - if you want to make a truly useful service out of Watson, they do it as a consultation and do it essentially from scratch.