Paper accepted to Computer Vision and Image Understanding