Fengwei Yu's home page

Research

My current work centers on the systems and infrastructure behind large-scale video AI — from massive pretraining to real-time generation and serving. Areas of interest:

• Ultra-large-scale video pretraining
• Low-latency real-time video generation and inference
• Real-time video director agents
• High-performance inference serving architecture
• Globally distributed video infrastructure
• Large-scale data platforms

News

• [2024] 1 paper accepted at IJCV.
• [2023] 5 papers accepted at CVPR, TPAMI, ICPP, MLSys, and IPDPS.
• [2022] 5 papers accepted at ICLR, CVPR, NeurIPS, and ICPP.
• [2021] 7 papers accepted at ICLR, CVPR, NeurIPS, and ICCV.
• [2021] Our team won the championship of the LPCVC 2021 FPGA Track, with our winning solution released as open source. We also released the MQBench quantization benchmark and the UP training framework.
• [2020] 4 papers accepted at ICPP, CVPR, and ICLR.
• [2019] 1 paper accepted at ICCV.
• [2016] We won 1st place in the ECCV MOT16 Challenge (data); 1 paper accepted at ECCV.

About me

In 2025 I co-founded an AI company building infrastructure for large-scale video AI; Series A at a $1.32B valuation.

Prior to that, I was Senior R&D Director at Momenta (HKEX: 6880), a leader in autonomous driving, where I led the training and inference infrastructure team — building the systems that took perception and planning models from research prototypes to on-vehicle deployment across a broad range of automotive chips.

Before Momenta, I was Vice Research Director at SenseTime (HKEX: 0020), advised by Prof. Xiaogang Wang and Dr. Junjie Yan. I led the Model-ToolChain team behind SenseTime's internal AI platform — the shared training, compression, and deployment stack that powered the majority of the company's smart-city and smart-vehicle products, scaling model production by roughly 2× each year.

Some of our work is open-sourced under the ModelTC organization on GitHub, and I write about it on Zhihu at 踢翻炼丹炉.

I first joined SenseTime as a full-time intern for two years in the Video Intelligence Group under Dr. Junjie Yan, working on high-performance video analysis systems (face, vehicle, pedestrian) and porting algorithms to embedded platforms; during that stretch we took 1st place at the ECCV MOT16 Multi-Object Tracking Challenge.

During college, I led the BUAA HPC team to first prize at the ASC Student Supercomputer Challenge 2015, and spent half a year interning on the Microsoft IoT team in Beijing, working on IoT communication protocols.

I earned my B.S. in 2015 and my M.S. in 2018, both from Beihang University CS. My M.S. was advised by Prof. Xiaohua Shi and spanned three separate directions: JVM memory profiling; CompCert (a formally verified compiler); and applying large-scale deep-learning language models to text error detection and correction.

I picked up competitive informatics in high school; back-to-back 1st Prizes at the National Olympiad in Informatics in Provinces (NOIp, ranked #55 nationally) recruited me into the Beihang University CS program.

Awards

• [2021] Champion, LPCVC 2021 FPGA Track (as team lead).
• [2017] Graduate National Scholarship, Beihang University.
• [2016] Huawei Scholarship, Beihang University.
• [2016] 1st place, ECCV MOT16 Multi-Object Tracking Challenge.
• [2015] 1st Prize (5th globally), ASC15 Student Supercomputer Challenge (team lead, Beihang HPC).
• [2010, 2009] 1st Prize (national rank #55), National Olympiad in Informatics in Provinces (NOIp).

Publications

2024

2023

2022

2021

2020

2019

2016