About Me

I obtained my Ph.D. in Computer Science from University of Rochester. I am very happy to announce that I will join Shanghai Jiao Tong University as a tenure-track assistant professor in John Hopcroft Center for Computer Science in Spring'24.

During my Ph.D., I worked with Prof. Yuhao Zhu who really shows me the intricacy and excitment of computer system and computer archtecture. My main research interest is about Mobile Visual Computing. This topic is not only about how machines perceive visual information, but also how machines present visual information to us! My current interests include but are not limited to in-sensor computing, point cloud-based computing and XR rendering. In my leisure time, I also play basketball and program for fun 🏀💻🏀💻.

Before I came to U of R, I obtained a Master degree from Carnegie Mellon University in Material Science. I worked with Prof. Rollett as a researcher on scientific computing.

Contact Details

Yu Feng
3002 Wegmans Hall
Rochester, NY 14620 US

(412)706-4492
yfeng28@ur.rochester.edu

Education

University of Rochester

Advisor: Yuhao Zhu
Ph.D. in Computer Science. Rochester, NY. 2017.09-2023.07.

Carnegie Mellon University

Advisor: Anthony Rollett
Master of Material Science. Pittsburgh, PA. 2015.09-2017.05

Tianjin Polytechnic University

Bachelor of Material Science. Tianjin, CHINA. 2011.09-2015.06

Work Experience

University of Rochester

Postdoc. September 2023 - Now

Meta Reality Labs

Foveated VR Passthrough. May 2022 - October 2022

Facebook Reality Labs

Event-based Eye Tracking. September 2020 - December 2020

Google AI

Smart Code Instruction Prefetching. May 2019 - September 2019

Publications

Conference Articles

Amanda: Unified Instrumentation Framework for Deep Neural Networks, ASPLOS 2024.

Yue Guan, Yuxian Qiu, Jingwen Leng, Fan Yang, Shuo Yu, Yunxin Liu, Yu Feng, Yuhao Zhu, Lidong Zhou, Yun Liang, Chen Zhang, Chao Li, Minyi Guo

[Invited Paper] Learned In-Sensor Visual Computing: From Compression to Eventification, ICCAD 2023.

Yu Feng, Tianrui Ma, Adith Boloor, Yuhao Zhu, Xuan Zhang

CamJ: Enabling System-Level Energy Modeling and Architectural Exploration for In-Sensor Visual Computing, ISCA 2023. [code] [doc] [tutorial]

Tianrui Ma*, Yu Feng*, Xuan Zhang, Yuhao Zhu (*: equal-contribution)

Fast and Accurate: Video Enhancement Using Sparse Depth, WACV 2023.

Yu Feng, Gunnar Hammonds, Yiming Gan, Yuhao Zhu

Crescent: Taming Memory Irregularities for Accelerating Deep Point Cloud Analytics, ISCA 2022. [code]

Yu Feng, Patrick Hansen, Paul Whatmough, Guoyu Lu, Yuhao Zhu

Real-Time Gaze Tracking with Event-Driven Eye Segmentation, IEEE VR 2022, Best Paper Honorable Mention.[slide][talk][code]

Yu Feng, Nathan Goulding-Hotta, Asif Khan, Hans Reyserhove, Yuhao Zhu

Mesorasi: Enabling Efficient Point Cloud Analytics via Delayed-Aggregation, MICRO 2020. [slide][talk][code]

Yu Feng*, Boyuan Tian*, Tiancheng Xu*, Paul Whatmough, Yuhao Zhu (*: equal-contribution)

Real-Time Spatio-Temporal LiDAR Point Cloud Compression, IROS 2020. [slide][talk][code]

Yu Feng, Shaoshan Liu, Yuhao Zhu

ASV: Accelerated Stereo Vision System, MICRO 2019. [slide]

Yu Feng, Paul Whatmough, Yuhao Zhu

PES: Proactive Event Scheduling for Responsive and Energy-Efficient Mobile Web Computing, ISCA 2019. [slide]

Yu Feng, Yuhao Zhu

Journal Articles

Extension of the Mechanical Threshold Stress Model to Static and Dynamic Strain Aging: Application to AA5754-O, 21 August 2017.

Yu Feng, Sudipto Mandal, Brian Gockel, Anthony D. Rollett

Research

Taming Memory Irregular in Deep Point Cloud Analytics

July 2021 - February 2022

Point cloud has become an important modality in many computer vision applications. However, the main operations in point cloud algorithms are memory-inefficient due to its irregularity and redundancy. We proposed a algorithm-hardware co-design approach to tame the memory irregularity without compromising the overall accuracy.

LiDAR-guided Video Enhancement

Match 2021 - July 2021

Pixel flow is an important cue for many video enhancement tasks. However, accurate pixel flow is hard to obtain for many monocular vision tasks. This work proposes a lightweight and fast algorithm to leverage the synchronized LiDAR to obtain a sparse pixel flow. Using the sparse pixel flow, we propose an unified framework to improve the performance of many video enhancement tasks.

Real-time Event-driven Eye Tracking for AR/VR

September 2020 - February 2021

Eye tracking is an key application to enable many AR/VR applications. This project is inspired by the novel event-camera and design a mechanism to emulate the events in software. The software-emulated event allows us to effectively predict the eye region and only process the useful pixels to speedup the eye tracking pipeline.

Efficient Deep Learning for Point Cloud

January 2020 - April 2020

Point cloud becomes a common modality in deep learning. We found some common computation patterns that are unique to point cloud, and designed a framework for point cloud deep learning system.

Real-time Point Cloud Compression

October 2019 - February 2020

Point cloud is a key modality in many visual applications, such as autonomous driving. To enable real-time communications and data transfer, efficient real-time point cloud compression is important. We leverage the temporal and spatial redundancies in and across point cloud and develop an efficient compression technique.

Co-Design Stereo Vision System

October 2018 - April 2019

The key for Stereo Vision applications is the ability to obtain theambient information and estimate the depth of their surround-ings. We designed a framework, Stereo Engine, that leverages the unique characteristics in stereo vision and supports a wide range of algorithms in this domain.

Task Scheduling in Runtime Application

March 2018 - November 2018

Web applications are mainly user-driven/user-oriented. When and what tasks will be triggered and executed largely depends on user behaviors. It is a open issue for mobile developers to design a scheduler that can fulfill user experience (to speedup) meanwhile decrease the amount of energy consumption. We proposed a proactive event scheduler. Instead of optimizing based on current system state, we expand our optimization scope to speculate future system state. By increasing our optimization scope, we expect to enhance the user experience and lower the energy consumption.

Skills

Programming Languages

C, C++, Python, Java, Halide, Ruby

Other Toolkits

Matlab, OpenGL(basic)

Database

MySQL, HBase(basic)