Exploiting Low-level Image Segmentation for Object Recognition