Văn bản này: Bottom-up top-down cues for weakly-supervised semantic segmentation