Short Bio

Yujun Shen is currently a fourth-year Ph.D. student in Multimedia Lab (MMLab), The Chinese University of Hong Kong (CUHK), supervised by Prof. Xiaoou Tang and Prof. Bolei Zhou. He received his bachelor’s degrees of Electronic Engineering and Management in Tsinghua University in 2016. His research interests mainly lie in Computer Vision, Deep Learning, Generative Model, Network Interpretation, Explainable Artificial Intelligence (XAI).

News
  • [2020/03] Our research group GenForce is online!
  • [2020/02] Two papers got accepted by CVPR 2020!
  • [2019/05] Join Facebook Reality Lab (FRL) as a research intern!
  • [2018/07] Join Google as a software engineer intern!
  • [2018/03] One paper got accepted by CVPR 2018!
Experiences
GenForce (Jan. 2019 - Present)
Primary Team Member
Facebook (May. 2019 - Sep. 2019)
Research Intern
Google (Jul. 2018 - Dec. 2018)
Software Engineer Intern
The Chinese University of Hong Kong (Aug. 2016 - Present)
Ph.D. Student in Department of Information Engineering
SenseTime (May 2015 - Jul. 2016)
Research Assistant
Tsinghua University (Aug. 2012 - Jul. 2016)
Bachelor's Degree in Electronic Engineering
Tsinghua University (Aug. 2013 - Jul. 2016)
Second Bachelor's Degree in Management
Projects
Generative Modeling
In-Domain GAN Inversion for Real Image Editing
This work raises a new problem in the GAN inversion task, which is that the inverted code should not only recover the target image from pixel values, but also semantically present the image, i.e., in-domain codes. For this propose, we propose to first train a domain-guided encoder to project an image back to the native latent space, and then perform domain-regularized optimization to refine the code for a better reconstruction. Such in-domain codes significantly facilitate real image editing with fixed GAN models.
[Paper] [Project Page] [Code] [Demo]
Image Processing Using Multi-Code GAN Prior
This work aims at applying pre-trained GAN models to real image processing. Multiple latent codes are employed to invert a given image, whose feature maps are then composed at some intermediate layer of the generator. Such over-parameterization significantly advances GAN inversion task and enables a variety of real-world applications by reusing the rich knowledge GAN models have learned.
[Paper] [Project Page] [Code]
Semantic Hierarchy Emerges in Deep Generative Representations for Scene Synthesis
This work conducts detailed analysis on the layer-wise generative representation of GANs for scene synthesis. It turns out that GANs spontaneously learn hierarchical semantics, including layout (bottom layers), categorical objects (middle layers), and scene attributes (top layers). A re-scoring technique is further proposed to identify the semantics most relevant to the pre-trained model from a broad candidate set.
[Paper] [Project Page] [Code] [Demo]
Interpreting the Latent Space of GANs for Face Editing
Given a well-trained unconditional face synthesis model, this work proposes InterFaceGAN, which is capable of controling the attributes of generated faces by interpreting the semantics hidden in the latent space.
[Paper] [Project Page] [Code] [Demo]
Internship at Facebook
Fisheye Image Analysis
This project aims at applying deep learning model on Fisheye images.
Internship at Google
Ads Image Analysis
This project focuses on predicting the saliency map of Ads images.
Model Compression
Residual Knowledge Distillation
This work proposes Residual Knowledge Distillation (RKD) by distilling the knowledge from teacher model twice. Besides the student model, an assistant model is introduced to learn the residual error between the student and the teacher for better performance.
[Paper]
Stage-by-Stage Knowledge Distillation
This work improves KD method by proposing Stage-by-Stage Knowledge Distillation (SSKD). In particular, SSKD introduces progressive training manner and does not rely on the loss weight for training. It shows strong superiority and robustness on various tasks, including image classification, image detection, and face recognition.
[Paper]
Identity-Preserving Face Synthesis
Identity-Preserving Face Synthesis via Generating Semantic Features
This work proposes FaceFeat-GAN to balance the trade-off between image quality and diversity in conditional GAN. Specially, it first generates semantic facial features as intermediate outputs with a collection of feature generators, and then maps these features to photo-realistic image. A novel hierarchical adversarial competition is proposed to help all components collaborate together.
[Paper]
Learning a Symmetry Three-Player GAN for Identity-Preserving Face Synthesis
This work proposes FaceID-GAN that achieves high-quality identity-preserving face synthesis by introducing an additional player to conventional GAN model. Specially, the generator competes with not only the discriminator but also the classifier. FaceID-GAN also involves information symmetry in the training process to make sure the knowledge fed into the generator is consistent with the supervision of generator.
[Paper] [Demo]
Honors and Awards
  • Hong Kong Ph.D. Fellowship, Hong Kong Research Grants Council, 2016
  • Excellent Graduate, Tsinghua University, 2016
  • Excellent Graduation Project, Tsinghua University, 2016
  • Academic Excellence Scholarship, Tsinghua University, 2014, 2015
  • Honorable Mention, The Mathematical Contest in Modeling (MCM), 2014
  • First prize, National Competition in Physics for College Students, 2013
  • National Scholarship, Ministry of Education of China, 2013
  • A member of "Project for Leadership", Tsinghua University, 2012
Skills
  • Programming Language: Python, C/C++
  • Deep Learning Framework: Tensorflow, PyTorch, Caffe
  • Tools: GitHub, LaTex, Office
  • Platforms: Windows, Linux, MaxOS
  • Language: Mandarin, English
  • Hobbies: Soccer, Badminton, Singing