Chunhui Liu

I'm a Senior Machine Learning Engineer at TikTok's Video Search Team, where I serve as the tech lead for relevance and pretraining. I work with a group of talented individuals to build the world's largest short video search engine with tens of billions videos, and serving billions of users daily. My team focuses on developing BERT/LLM models as part of the search engine to ensure that the results are safe and relevant, utilizing advanced techniques in CV/NLP/Multimodal learning and pretraining.

I was previously an Applied Scientist at Amazon AI, where I conducted cutting-edge research and developed real-world applications for video and action understanding from cameras. I obtained my Master's degree in Computer Vision from CMU and my Bachelor's degree in Computer Science, Summa Cum Laude, from Peking University, under the supervision of Prof. Jiaying Liu.

Email  /  Scholar  /  Github

We are hiring!

Research and Selected Publications

Recently, I have not been an active researcher on public topics. However, I remain deeply interested in deep learning, computer vision, multi-modality learning, and video understanding. I am passionate about building machine learning models that understand the physical world in a way similar to humans.

The first SOTA transformer model on Action Detections, using learnable queries as tubelet proposals.

Utilizing transformers as spatial feature sampler, achieve 6x faster inference without performance drop.

One of the earliest works to use the transformer architecture for action recognition

We present a survey paper that summarizes 16 datasets and 200 existing papers on action understanding. Additionally, we provide tutorial workshops and a complete codebase to help newcomers join this field.

A skeleton-baded action detection dataset.

This template is forked from source code.