Learning domain-specific cameras