Coming soon:
Face dataset with ambiguous labels from screenplay, as in Learning from Ambiguously Labeled Images.