Summary of the 2024 Low-Power Efficient and Accurate Facial-Landmark Detection for Embedded Systems

Yu-Shu Ni, Han-Chun Chen, Chia-Chi Tsai, Chih-Cheng Chen, Po-Yu Chen, Hsien-Kai Kuo, Jun-Ying Hunag, Po-Chi Hu, Jenq-Neng Hwang, Jiun-In Guo

Abstract: The 2024 IEEE ICME Grand Challenge has been a pivotal event in advancing low-power efficient and accurate facial-landmark detection for embedded systems. It emphasized the development of models that are both energy-efficient and highly accurate. With participation from 165 teams, the competition provided an extensive dataset for model training, highlighting facial diversity to bolster model resilience. This competition unfolded in two phases: an initial online evaluation and a subsequent final round conducted on the MediaTek Dimensity 9300 Series platform, aimed at assessing the efficacy of real-time applications. A critical aspect of the competition was the models' ability to precisely detect 51 facial landmarks, a feature integral to applications such as autonomous driving, by capturing detailed facial expressions. The victors of the challenge were recognized based on their models' accuracy and minimal power consumption, marking a notable advancement in the field of facial-landmark detection and significantly advancing applications in embedded systems.