Junyi Cao
Master Student
CSE @ SJTU
Hard work does not always pay off.
Yet that is the reason why persistence is precious.
Yet that is the reason why persistence is precious.
About Me
I recently graduated with a Master’s Degree in computer science and technology at the Department of Computer Science and Engineering, Shanghai Jiao Tong University, supervised by Prof. Chao Ma. Previously, I obtained my Bachelor’s Degree in software engineering from South China University of Technology, advised by Prof. Mingkui Tan.
My research focuses on building intelligent machine systems that are able to (1) sense, (2) understand, and (3) reason in the real world like humans. In pursuit of this goal, I have conducted research on
- deepfake & spoofing detection,
- scene reconstruction & understanding,
- visual dynamics grounding.
Recently, I have taken a step further and am focusing more on AI systems’ interaction with the physical world.
News
- [Sep. 2024] One first-author paper is accepted by NeurIPS 2024. Please see this page for more details.
- [Aug. 2024] Began a new internship at TUC supervised by Prof. Evangelos Kalogerakis for 4D generation.
- [Mar. 2024] Graduated with honer from SJTU.
- [Feb. 2024] Began a new internship at vivo researching 3D representations for generative dynamics.
Publications
(* indicates equal contribution)NeuMA: Neural Material Adaptor for Visual Grounding of Intrinsic Dynamics
Junyi Cao,
Shanyan Guan,
Yanhao Ge,
Wei Li,
Xiaokang Yang,
Chao Ma
In
Advances in Neural Information Processing Systems (NeurIPS),
2024.
Key points: We present an optimization-based method that integrates widely accepted physics laws with learnable corrections for grounding intrinsic dynamics from visual observations. The learned dynamics can be applied to new environments and used for physically plausible 4D generation.
Lightning NeRF: Efficient Hybrid Scene Representation for Autonomous Driving
Junyi Cao,
Zhichao Li,
Naiyan Wang,
Chao Ma
In
IEEE International Conference on Robotics and Automation (ICRA),
2024.
Key points: We propose an efficient hybrid scene representation for autonomous driving scenarios that reduces time complexity for training and rendering via the geometry prior from LiDAR observations.
Towards Unified Defense for Face Forgery and Spoofing Attacks via Dual Space Reconstruction Learning
Junyi Cao,
Ke-Yue Zhang,
Taiping Yao,
Shouhong Ding,
Xiaokang Yang,
Chao Ma
In
International Journal of Computer Vision (IJCV),
2024.
Key points: We put forward a dual space reconstruction learning framework that focused on the commonalities of real faces in both spatial and frequency domains to learn the comprehensive difference between real faces and diverse attacks. In addition, we set up a novel benchmark consisting of both face forgery attacks and face spoofing attacks to evaluate models’ competence against diverse attack data.
End-to-End Reconstruction-Classification Learning for Face Forgery Detection
Junyi Cao,
Chao Ma,
Taiping Yao,
Shen Chen,
Shouhong Ding,
Xiaokang Yang
In
IEEE International Conference on Computer Vision and Pattern Recognition (CVPR),
2022.
Key points: We train a reconstruction network over genuine faces only and use the output of the latent feature by the encoder to perform binary classification.
Dynamic Extension Nets for Few-shot Semantic Segmentation
Lizhao Liu*,
Junyi Cao*,
Minqian Liu*,
Yong Guo*,
Qi Chen*,
Mingkui Tan
In
ACM International Conference on Multimedia,
2020.
Key points: We build dynamic extension nets for few-shot segmentation which constructs and maintains a classifier for the novel semantic class by leveraging the knowledge from the base classes.
