Abstract:In this paper, we proposed a new method called Prior Landmark Algorithm (PLA) to address the limitations of traditional face landmark prediction techniques. PLA predicts the offset between real and prior landmarks to overcome the issue of solvable domain coverage and to handle different face poses and shapes. A new loss function with weights (PL-Loss) and the strategy of Online Hard Example Mining were also introduced to improve the accuracy of the model. The adoption of prior landmarks and an auxiliary classifier eliminate the need for face normalization during training. The results show that PLA achieved comparable accuracy to state-of-the-art models and had a small model size of only 10 MB with fast processing speed of 162 FPS on average.

PLA: Fast and Accurate Face Alignment Network based on Prior Landmarks