Revising FUNSD dataset for key-value detection in document images