PyraMoT: A Novel Framework for Enhanced Facial Thermal Landmarks Detection

Facial analysis and recognition are vital components in many real-world applications, such as driver safety monitoring, security systems, and healthcare. Despite the need for thermal images in many of these applications, there has been limited research conducted on facial landmark recognition for thermal images. This scarcity can be partly attributed to a lack of thermal datasets for comprehensive analysis. In this paper, we make three major contributions to the field of thermal facial landmark detection. First, we present D5050, a novel thermal face dataset that includes 166 video sequences and 5050 annotated thermal images from 163 different subjects, which is, to our knowledge, the largest dataset with comprehensive landmark annotation in terms of the number of subjects. Second, we propose PyraMoT, an innovative thermal facial landmark detection framework that combines a customized encoder-decoder structure and a Feature Pyramid Network (FPN) with the efficiency of MobileNetV2and the incorporation of the anisotropic loss function. Third, we conduct a thorough comparative analysis with existing approaches, including six methods, three datasets, and four color palettes for thermal facial landmark detection. This study employs measures such as Normalized Mean Error (NME) and Failure Rate (FR) to determine and comparatively evaluate the accuracy and reliability of the various detection methods. Overall, our proposed approach outperforms other methods on different datasets and contributes to the field of thermal image processing, proposing three main advancements that address current limitations in the field.