Decoding what people see from where they look: Predicting visual stimuli from scanpaths