We investigate the importance of parts for the tasks of action and attribute
classification. We develop a part-based approach by leveraging convolutional
network features inspired by recent advances in computer vision. Our part
detectors are a deep version of poselets and capture parts