Haodong Duan
Research Scientist at ByteDance
Shanghai, Hong Kong
Summary
🤩
Rockstar🎓
Top SchoolHaodong Duan is a research scientist with eight years of experience bridging academic excellence and production-grade open-source engineering in multimodal AI. He holds a Ph.D. from CUHK and a B.S. from Peking University, has published 30+ papers with 6,500+ citations, and led the creation of MMBench and the Open VLM Leaderboard—now key fixtures in multimodal evaluation. At Shanghai AI Lab he led the multi-modal evaluation team and contributed core components to widely used toolkits like MMAction2, PySkl, MMCV and OpenCompass, reflecting deep expertise in video understanding and evaluation infrastructure. His work combines rigorous benchmarking, dataset and feature-pipeline engineering (e.g., flow/gen tools and skeleton-based action modules), and practical improvements to LLM evaluation tooling. Now at ByteDance, he continues to translate cutting-edge research into widely adopted open-source systems and authoritative leaderboards.
8 years of coding experience
High School Diploma, High School Diploma at No.2 High School of East China Normal UniversityBachelor of Science - BS, Computer Science, Bachelor of Science - BS, Computer Science at Peking UniversityThe Chinese University of Hong Kong (CUHK)
English, Chinese
Github Skills (30)
pytorch10
benchmark10
python10
evaluation10
video-processing10
machine-learning10
benchmarking10
video-understanding10
ml10
llm10
mle10
eval10
deep-learning10
action-recognition10
sd-mmc10
Programming languages (8)
Github contributions (5)
kennymckormick/pyskl
Apr 2022 - Jan 2023
A toolbox for skeleton-based action recognition.
Role in this project:
Back-end Developer & ML Engineer
Back-end Developer & ML EngineerContributions:1 release, 5 reviews, 91 commits in 9 months
Contributions summary:Haodong primarily worked on the core functionality of the `pyskl` repository, focusing on skeleton-based action recognition. The commits demonstrate modifications to the `smp.py` file, likely involving updates to the skeleton processing or model training pipeline. Several commits involve adding and modifying the `custom_2d_skeleton.py` file, suggesting contributions to a custom 2D skeleton extraction tool. Moreover, the user added configurations and examples related to action recognition tasks, including support for the Diving48 dataset.
stgcnskeletondeep-learningposec3dskeleton-based-action-recognition
open-mmlab/mmaction2
Jun 2020 - Jun 2022
OpenMMLab's Next Generation Video Understanding Toolbox and Benchmark
Role in this project:Back-end Developer & ML Engineer
Back-end Developer & ML EngineerContributions:12 releases, 426 reviews, 367 commits in 1 year 11 months
Contributions summary:Haodong's commits primarily focused on implementing and modifying functionality related to flow estimation and processing within the mmaction2 repository. They added a `gen_flow` tool with support for different algorithms, which suggests involvement in developing or integrating features for video analysis. Further modifications to existing code, including changes to data loading pipelines and models for various action recognition tasks (TSN, SlowFast, etc.), demonstrate their role as an engineer contributing to the project's core functionality. Additionally, the introduction of a feature extraction tool indicates a focus on preparing or processing data for training or analysis.
avax3dvisual-recognitionbenchmarkvideo
Find and Hire Top DevelopersWe’ve analyzed the programming source code of over 60 million software developers on GitHub and scored them by 50,000 skills. Sign-up on Prog,AI to search for software developers.