Main Projects
Aorta: Image Registration and Deformation Analysis (Ongoing...)
Goal: Measure the deformable change of aorta along the time.
Achievement: Propose a novel method to decompose the Jacobian into radial and longitudinal directions; Apply the DL-based segmentation on CT; Automate the centerline extraction.
Keywords: Registration, Directional Jacobian, CNN for Segmentation
Weakly-supervised Segmentation (Published)
Goal: Perform weak-supervised segmentation for skin lesion images (vitiligo).
Achievement: Propose a dataset; Propose a weak-supervised segmentation for vitiligo images, and achieves IoU at 71.4 which is comparable with SOTA ( e.g. PRM, SEC-CRF.)
Keywords: CNN, Saliency Propagation, Weakly-supervised Segmentation
Highlight Removal in Facial Images (In submission)
Goal: Remove specular highlight in facial images
Achievement: Our method generates highlight-free images with high-quality details and fewer artifacts by adopting CGAN and multi-scale discriminators.
Keywords: Highlight Removal, Facial Image, CGAN
Estimate 6D Pose from Single RGB Image with Backpropagatable PnP
Goal: Estimate 6D pose from single RGB image.
Achievement: Our method achieves BOP score at 61.4 on YCB-V using only RGB images, which is 10.6 higher than original Pix2Pose’s performance.
Keywords: 6D pose, PnP, U-net, single RGB
>>> See more results: [report]
Video Synopsis: Keyframe Extraction and video condensation
Goal: Apply key frame extraction and video condensation to surveillance videos.
Achievement: Implement a video condensation system from scratch using C++; Designed a novel motion-volume based keyframe extraction method; Second Prize of National-level SRTP(Student Research Training Program)
Keyword: K-means, Motion-volume, Graph-cut, Tracking ...
Other Projects
Achievement: Greatly improved the tracking robustness for planar object in the AR application, compared to AR-Toolkit framework; Developed an AR game based on improved framework.
Keywords: AR, Planar Object Tracking, Graph Match ...
Achievement: Implemented Double/Dueling DQN algorithm and Prioritised Experience method from scratch using Pytorch; and achieved, even exceeded a bit, the performance reported by original papers.
Keywords: Reinforcement Learning, Atari, Deep Q Learning ...
>>> See more results: [Github repo]