Part-aware distillation and aggregation network for human parsing

Abstract

The current state-of-the-art human parsing models achieve remarkable success in parsing accuracy. However, the huge model size and computational cost restrict their applications on low-latency online systems or resource-limited mobile devices. In this paper, we propose a novel part-aware distillation and aggregation network for human parsing, which can be applied to any human parsing model to achieve a good trade-off between accuracy and efficiency. We design the part key-point similarity distillation and the part distribution distillation to transfer the complex teacher model’s knowledge of part structural and spatial relationships to the lightweight student model, which can help the latter to better identify small parts and semantic boundaries, and to distinguish easily confused categories. Furthermore, the online model aggregation module is introduced in the later stages of training, which can mitigate noise from both the teacher and the labels to obtain smoother and more robust results. Extensive experiments and ablation studies on the large-scale popular human parsing datasets LIP, ATR and PASCAL-Person Part fully demonstrate that our method is accurate, lightweight and general.