A basic point about assessing recognisers is that you cannot assess how well a recogniser has performed unless it has been trained with data whose reliability are known. Thus, the issues of assessing human segmentation and classification performance have to be addressed.