How to learn pixel-level representations of human parts without
supervision is a challenging task. However, despite its significance, a
few works explore this challenge. In this work, we propose a contrastive
and consistent learning network (C2L) for unsupervised human parsing.
C2L mainly consists of a part contrastive module and a pixel consistent
module. We design a part contrastive module to distinguish the
same semantic human parts from other ones by contrastive learning,
which pulls the same semantic parts closer and pushes different semantic
ones away. A pixel consistent module is proposed to obtain spatial correspondence in each view of images, which can select semantic-relevant image pixels and suppress semantic-irrelevant ones. To improve the pattern analysis ability, we perform a sparse operation on the feed-forward networks of the pixel consistent module. Extensive experiments on the popular human parsing benchmark show that our method achieves competitive performance.