Using self-context for multimodal detection of head nods in face-to-face interactions