Spatial Gradient Guided Learning and Semantic Relation Transfer for Facial Landmark Detection

Jian Wang, Yaoyi Li, Hongtao Lu

Abstract: Pixel-wise losses are widely used in heatmap regression networks to detect facial landmarks, however, those losses are not consistent with the evaluation criteria in testing, which is evaluating the error between the highest pixel position in the predicted heatmap and the ground-truth heatmap. In this paper, we proposed a novel spatial-gradient consistency loss function (called Grad loss), which maintains a similar spatial structure in the heatmap with ground-truth. To reduce the quantization error caused by downsampling in the network, we also propose a new post-processing strategy based on the Gaussian prior. To further improve face alignment accuracy, we introduce Spatial-Gradient Enhance attention and Relation-based Reweighing Module to transfer semantic information and spatial information between high-resolution and low-resolution representations. Extensive experiments on several benchmarks (e.g., 300W, AFLW, COFW, WFLW) show that our method outperforms the state-of-the-art by impressive margins.