Develop a visual cue extraction framework to enhance the ability of a nurse robot (Grace) to interact and engage with patients during interviews. Various visual perception understanding tasks are explored including Object Detection, Tracking, and Reidentification, 2D-3D Body/Head Pose Estimation, Facial Expression Recognition and Analysis, Abnormal Action Recognition...

Propose a method for generating highly realistic medical cancer images given any pairs of non-tumorous background and tumour object images, directly utilizing the pretrained Stable Diffusion.

Attempt to improve FML by quantifying the uncertainty levels of the meme and local models for each client and leverage it as the weighting scheme for both global model aggregation and client update stages.