说明：  This paper presents a novel yet intuitive approach to unsupervised feature learning. Inspired by the human visual
(effective visual representation Specifically we use unsupervised motion-based segmentation on videos to obtain segments)