Part-aware distillation and aggregation network for human parsing

Yuntian Lai, Yuxin Feng, Fan Zhou, Zhuo Su

Abstract: Highlights•We introduce knowledge distillation into the human parsing task, proposing a part-aware distillation and aggregation network for human parsing.•The part key-point similarity distillation and the part distribution distillation are designed to transfer part structural and spatial relationships from the teacher to the student.•We propose the online model aggregation module to mitigate noise from both the teacher and the labels, and to prevent overfitting in the student during the later stages of training.