Published on 17 Jun 2022

SCSE Outstanding PhD Thesis Award 2022

Congratulations to the following PhD graduates for their achievement!

Winner 
for contributions to research in generative image modeling and synthesis, specifically on topics of image completion, image translation and scene decomposition.

Passport size photo of awardee - Dr Zheng Chuanxia.

Dr ZHENG Chuanxia 
Synthesizing Photorealistic Images with Deep Generative Learning

Citation: 
The research presented in Zheng Chuanxia’s PhD thesis deals with generative image modeling and synthesis. One of the contributions made is “pluralistic image completion”, a mathematically principled approach to completing partial images focused on preserving the diversity of possible solutions; it is rapidly becoming a standard benchmark for other competing methods, and has since attracted interest in commercial licensing from three global companies. Another of Chuanxia’s contribution is a new spatial correlative loss that can be generalized to many unpaired image translation tasks, anchored on a fundamental insight that congruence of scene structure between images, especially in very different domains, is best captured through spatial patterns of self-similarity. Other notable contributions in this thesis include task-mediated image translation for depth estimation, and a “visit-the-invisible” framework for scene decomposition and completion. 



Runner-Up 
for contributions to visual captioners for generating more descriptive and less biased captions.

Passport size photo of awardee - Dr Yang Xu.

Dr YANG Xu 
Incorporating Additional Knowledge into Image Captioners

Citation: 
Although modern deep learning-based captioners can generate fluent captions to describe the given images, these captions are usually less descriptive and easily affected by the dataset bias. Yang Xu’s thesis focused on alleviating these two limitations with the following significant contributions: (1) He proposed to exploit scene graphs to transfer language inductive bias knowledge from pure language domain to vision-language domain for more descriptive captions. (2) He built a module network-based captioner for generating less biased captions. (3) More importantly, he retrospected previously proposed captioners, including his, from the perspective of causal inference and then developed a principled solution for eliminating dataset biases. (4) He also introduced a visual feature pre-training strategy for detecting less object-biased relationships and a scene graph-based paragraph generator for more fine-grained paragraphs. 



Jury 2022

Assoc Prof Adams Kong (Jury Chair)

Assoc Prof Sinno Pan

Asst Prof Li Boyang