LAUSR

Abstract In this paper, we propose a geometry-attentive relational reasoning approach to investigate the problem of robust facial landmark detection, especially when faces were captured in wild conditions. Unlike existing methods which usually cannot explicitly exploit the geometric relationship among different landmarks, our approach aims to reason about the intrinsic geometry-aware relations among landmarks for feature enhancement. To achieve this, we carefully develop an interpretable and plug-and-play module to reinforce the discriminativeness and uniqueness of feature maps, which typically operates on all possible pairs on the immediate inter-landmark heat maps. Among these pairing maps, our model learns to infer the meaningful relational clues in the transformed feature space on condition of holistic facial shape prior. For permutation order invariance, we pool these features as a single aggregated relational feature. To further improve the performance, we simply equip the proposed module inside the backbone hourglass networks. The experimental results on the standard benchmarking datasets indicate the effectiveness of our proposed approach.