Chinese researchers develop a new model for 3D facial landmark detection

Chinese researchers say they have built a high-precision 3D facial database of about 200,000 scans and used it to develop a new facial landmark detection model aimed at improving humanoid robots and digital humans.

The scientists’ work was published in IEEE Transactions on Circuits and Systems for Video Technology.

According to the institute, the team led by Song Zhan of the Shenzhen Institutes of Advanced Technology of the Chinese Academy of Sciences and Ye Yuping of Fujian University of Technology, developed what it calls a curvature-fused graph attention network, or CF-GAT, that can predict facial landmarks directly from raw 3D point clouds.

Point clouds are sets of data points in 3D space, and in this case they are being used to represent facial geometry without relying on conventional 2D image textures or template-based models.

The researchers said existing 3D facial landmark detection methods have often been constrained by a lack of large, accurately annotated 3D face datasets, pushing many systems to depend on 2D texture assistance or synthetic 3D faces.

Their new method, they said, uses a geometry-driven sampling strategy to preserve curvature information and feed it into the network’s attention mechanism, allowing the model to capture local shape variations and broader facial structure from unordered point clouds.

The database assembled for the project includes roughly 200,000 high-fidelity 3D facial scans, along with related datasets covering multiple facial expressions, standardized 3D facial landmarks, human body data and dynamic 4D facial expressions.

The institute said the dataset was selected for Fujian Province’s 2025 High-Quality AI Dataset Program.

The Chinese Academy of Sciences framed the project as a step toward more lifelike robots and virtual humans. It said 3D facial keypoint detection is important for machines that need to express emotions, recognize identities and interact in more humanlike ways.

The researchers also reported that the model showed stronger robustness to noise, better generalization across different facial shapes and more accurate localization of fine-grained landmarks in testing.

The work also reflects a broader push in China to use point clouds in facial analysis. In a separate 2024 study, researchers from Peking University developed a deep learning model for age estimation using non-registered 3D face point clouds.

That work focused less on robotics and more on privacy, proposing coordinate-wise monotonic transformations that could mask facial data while preserving age-related features for machine analysis.

The Peking University-led team trained its system on more than 16,000 examples of 3D face point cloud data and reported an average absolute age estimation error of about 2.5 years.

The researchers argued that the approach could help reduce privacy risks by making faces harder for humans to identify while still allowing machines to extract age-related information.

Taken together, the two projects show how point cloud-based facial analysis is expanding in China across different applications.

One line of research is aimed at improving machine perception and facial realism in robots and digital humans, while another is exploring whether 3D facial geometry can support analytics such as age estimation with less exposure of directly identifiable face data.

The new 3D face database claim has drawn attention because of its size, though the figure comes from the project’s own announcement and has not been independently verified in the materials reviewed here.

Still, the project suggests that high-volume 3D facial data collection and analysis are becoming a more important part of biometric and AI research, particularly as developers look beyond flat images and toward richer geometric models of the human face.

Article Topics

3D point clouds  |  biometric dataset  |  biometrics  |  biometrics research  |  China  |  face biometrics  |  face detection  |  IEEE