Multi-scale Densely U-Nets Refine Network for Face Alignment

Jun Yu, Haonian Xie, Guochen Xie, Mengyan Li, Zengfu Wang

Abstract: We develop a Multi-Scale Densely U-Nets Refine Network (MSDUR-Net) for face alignment. This method improves the recent deep conv-deconv hourglass models with four key aspects: (1) Multi-scale supervision to enhance contextual feature learning by constraining feature heatmaps in multiple scales. (2) Refine Regression Network at the end to globally optimize the facial landmark by fusing the multi-scale features. (3) Densely U-Nets as the backbone instead of traditional hourglass network to improve information fiow and mitigate the information loss. (4) Integral regression as regression loss to reduce quantization errors. The experiments demonstrate that our model can handle many problems, such as scale varieties, large pose variation and partial occlusions effectively, which many popular methods always suffer. Our approach achieves the 2nd place in the Grand Challenge of 106-p Facial Landmark Localization in ICME 2019.