How Face Keypoint Localization Advances Under Masked Conditions: Insights from JD AI’s 3rd Competition

The JD AI Institute, in partnership with ICME2021, hosted the third Face Keypoint Localization Competition, aiming to discover efficient methods for locating facial landmarks when subjects wear masks. The technology supports tracking close contacts of COVID‑19 patients and broader epidemic control.

What is Face Keypoint Localization?

Face keypoint localization, also known as facial landmark detection, involves identifying crucial points on a face image, such as contours and facial features. It underpins applications like 3D face reconstruction, expression transfer, and special‑effect cameras.

The competition attracted 83 renowned universities and enterprises worldwide. After verification, model submission, and testing phases, the top five teams were from Meituan, Tencent, Streamax, ByteDance, and the University of Chinese Academy of Sciences.

This series has now run for three editions, with over 400 teams participating, including elite institutions such as Tsinghua University, Peking University, Wuhan University, Hong Kong University, Rochester Institute of Technology, National University of Singapore, and the University of Michigan, as well as major internet companies like Baidu, Sogou, Meituan, and Tencent.

Compared with previous editions, this year’s competition increased data scale and difficulty. The test set contains both real‑mask and synthetic‑mask images, and stricter limits on computational efficiency and model size were imposed, making the challenge more realistic and impactful.

The event not only fostered knowledge and skill development in artificial intelligence and computer vision for students and scholars but also generated innovative solutions and cultivated professional talent. JD AI Institute plans to continue leveraging its technology platform to empower various industries with AI, delivering more efficient and convenient services.