We consider a binary labelling problem: for some machine learning applications, two types of distinct objects are required to be labeled respectively, before a classifier can be trained. We show that the famous ESP game and variants would not work well on this binary labelling problem. We discuss how to design a new human computation game to solve this problem. It turns out that interesting but subtle security issues emerge in the new game. We introduce novel gaming mechanisms, such as ‘guess disagreement’, which improve the game’s security, usability and productivity simultaneously.