Learning and generalising semantic knowledge from object scenes