Towards Hardware-Friendly and Robust Facial Landmark Detection Method

Liang Xie, Menghao Hu, Xinbei Bai, Wenke Huang

Abstract: Facial Landmark Detection (FLD) plays an essential role in computer vision because it is the premise of many tasks such as face recognition and facial expression analysis. Although significant advancements have been achieved with the help of deep learning, the performance of FLD is still unsatisfactory due to the influence of occlusion, low illumination, and motion blur. Existing works are developed and implemented based on expensive computing GPUs, limiting their application. This paper proposes a hardware-friendly, fast, and high-performance FLD framework. We first utilize a lightweight CNN to extract its features given the face image. This procedure uses a multi-scale feature fusion strategy for better feature representation learning. We design a weighted model to guide the regression of other landmarks inspired by the spatial distribution of five key points on the face: the eyes, nose and mouth. Our proposed network can also be quantified and pruned for practical deployment running at 45 FPS on the ARM3288 chip. We collect and annotate a new dataset CTLM-100K, which contains 100K facial samples with various postures and lighting conditions. Extensive experiments on these three benchmark datasets all validated the effectiveness of our model.

External IDs:dblp:conf/iconip/XieHBH22