Fard Ali Pourramezan ; Mahoor Mohammad H.

Abstract

Although deep neural networks have achieved reasonable accuracy in solvingface alignment, it is still a challenging task, specifically when we deal withfacial images, under occlusion, or extreme head poses. Heatmap-based Regression(HBR) and Coordinate-based Regression (CBR) are among the two mainly usedmethods for face alignment. CBR methods require less computer memory, thoughtheir performance is less than HBR methods. In this paper, we propose anAdaptive Coordinate-based Regression (ACR) loss to improve the accuracy of CBRfor face alignment. Inspired by the Active Shape Model (ASM), we generateSmooth-Face objects, a set of facial landmark points with less variationscompared to the ground truth landmark points. We then introduce a method toestimate the level of difficulty in predicting each landmark point for thenetwork by comparing the distribution of the ground truth landmark points andthe corresponding Smooth-Face objects. Our proposed ACR Loss can adaptivelymodify its curvature and the influence of the loss based on the difficultylevel of predicting each landmark point in a face. Accordingly, the ACR Lossguides the network toward challenging points than easier points, which improvesthe accuracy of the face alignment task. Our extensive evaluation shows thecapabilities of the proposed ACR Loss in predicting facial landmark points invarious facial images.