“He is the most exceptional undergraduate
student I have seen over the past 17 years in the Hong Kong University of
Science and Technology (HKUST),” professor Chi Keung Tang said highly of
Qifeng Chen.
His research areas are diversified across
multiple aspects of computer vision, including intrinsic image decomposition,
stereo reconstruction, Markov Random Field (MRF) optimization, and optical flow
estimation. Multiple papers were published on and selected for full oral
presentations at ICCV and CVPR.
One of his creative innovations was
applying MRF optimization to nonrigid registration of 3D surfaces for the first
time. The resulting algorithm outperformed the state-of-the-art existing model and increased
accuracy by 3 times. He basically overturned a decade of work on nonrigid
registration and reset research in this area.
Most recently, Chen became an assistant
professor at HKUST and moved to deep learning and image processing. He has been
working on many innovative research topics, primarily focusing on how to
revolutionize the image processing pipeline with deep learning techniques.
He is also working on an emerging research
topic on photographic image synthesis from semantic layouts. For example, given
a semantic layout of a scene, can an AI system synthesize an image that depicts
the scene and output a photograph? Does AI have imagination? Can AI create
animation autonomously?
"The answer is
yes, as my research shows," Chen said confidently. He invented a
semi-parametric approach based on Cascaded Refinement Networks to perform
photographic image synthesis.
Chen is currently working on the
development of an AI tool that aims to improve the authenticity of film visual
contents and effects. This can reduce the cost of a movie by reducing the
amount of manual work for creating such contents and effects. His goal is to
upgrade this AI tool to automatically generate certain movie scenes or
characters based on high-level descriptions such as written scripts or user
scribbles.