Attentive occlusion-adaptive deep network for facial landmark detection

Abstract

To be very specific in this paper, an Attentive Occlusion-adaptive Deep Network, hereafter referred as AODN, is proposed for facial landmark detection, consisting of the geometry-aware module, attention module, and low-rank learning module. Facial Landmark Detection (FLD) is a fundamental pre-processing step of facial related tasks. Occlusion, extreme pose, different expressions and illumination are the main challenges in facial landmark detection related tasks. Convolutional Neural Network (CNN) based FLD methods have attained significant improvement regarding accurate FLD but, to deal with occlusion is still very challenging even for CNN. It is because; probably occlusion misleads CNN on feature representation learning. If faces are partially occluded, the localization accuracy will drop significantly. The role of attention in the human visual system is vital, and researchers proved its significance for the computer vision problem. Taking advantage of geometric relationships among different facial components and attention, we extended our already established Occlusion-adaptive Deep Network (ODN). We introduced the attention module consisting of Channel-wise Attention (CA) and Spatial Attention (SA) to improve its ability to deal with the occlusion and enhance feature representation ability simultaneously. The occlusion probability assists as adaptive weights of high-level features and minimizes the effect of the occlusion and assist in modelling the occlusion. Ablation studies prove the synergistic effect of each module. The summary of our trifold contribution is as follows: i) we introduced attention mechanism in our already established ODN model, to deal with occlusion more precisely, and get the rich feature representation to achieve better performance. ii) As per our best of knowledge, we are the pioneers to introduce CA and SA for FLD to model occlusion. iii) Our proposed methodology reduces the number of entire network parameters, which effectually decreases training time and cost. So, the proposed model is more suitable for scalable data processing. Experimental results prove the better performance of proposed AODN on challenging benchmark datasets.