Abstract Human body part parsing,or human semantic part seg- menianion.is fuindamental to many compuler vision tasks. In conventional semantic segmentntion methods,the ground rruth segmentations are provided,and fiully comvolutional nenorks (FCN)are lrained in an end-lo-end scheme.Al- though these methods hare demonstrated impressire result- s,their performance highly depends on the quantiry and qualiny of training data.In this paper we present a novel miethiod to generale svnthieric human part segmentarion da- ta using easily-obtained human keypoint annotations.Our key idea is to erploit the anatomical similarin among huu- man to transfer the parsing resulis of a person to another person with similar pose.Using these estimated results as additional training data.our semi-supenised model outper- forms its strong-supenised connterpart by 6 mlOU on the PASCAL-Person-Part dataset /dl.and we achieve state-of- the-art human parsing results.Our approach is general and can be readily extended to other objectlanimal parsing task assuming thar their anatomical similarity can be annotar- ed by kevpoint.s.The proposed model and accompanving source code will be made publicly available