Xiaojuan Qi

Xiaojuan Qi’s research focuses on advancing the frontiers of AI in 3D vision, with the goal of equipping AI systems with advanced spatial intelligence that enables them to perceive, understand, and interact with the physical world in human-like ways.

She pioneered the use of graph neural networks (GNNs) for 3D point cloud processing in RGB-D semantic segmentation, and her GeoNet series on monocular geometric estimation incorporates physical geometric constraints to produce more structurally consistent 3D reconstructions from single 2D images.

Since joining the University of Hong Kong, she has focused on open-world 3D scenarios, proposing hybrid neural representations for high-fidelity neural rendering and surface reconstruction, and introducing sparse-controlled Gaussian splatting (SC-GS) for dynamic 4D scene modeling. Her work on 3D object texture generation received a SIGGRAPH Asia Best Paper Honorable Mention. In 3D understanding, she developed advanced 3D processing architectures and a 3D self-training method, and explored knowledge transfer from pre-trained 2D models to 3D, reducing annotation data dependency.

Her research also extends to medical and biological applications, contributing to improved tumor diagnosis, advancements in electron microscopy 3D imaging, and the development of a deep learning–based physical unclonable function (PUF) anti-counterfeiting system.