They tested correlation between different perspectives (same scene and AP even) later in the paper and achieved an accuracy of 0%. Not to discount other methods being able to achieve that.