That is exactly what study 2 of this paper looks at and the answer is that if you control for age, gender and ethnicity, individual raters can't do much better

That is exactly what study 2 of this paper looks at and the answer is that if you control for age, gender and ethnicity, individual raters can't do much better that guessing (r=.02) but if you aggregate the ratings you start to gain accuracy (which matches results for aggregate guessing in other areas.)