2023 AI Training Dataset Construction Project
Heading a collaborative project with multiple companies to construct a comprehensive Hand-Object Interaction dataset.
From March to December 2023, I led a national AI training data initiative organized by the National Information Society Agency of Korea (NIA). This ambitious project focused on building a world-class dataset for hand-object interaction, with contributions from three industry partners and KAIST.
Project Scope and Collaboration
This was a comprehensive end-to-end R&D project encompassing:
- Data collection using a multi-view camera system
- Data annotation with custom-built toolkits and frameworks
- Rigorous verification and validation, including AI model development
As the project lead, I guided the overall planning and execution, directly implemented core processing methods, and coordinated collaboration across all partners—from technical roles to data sharing protocols.
Technical Approach & Implementation
Key accomplishments include:
- Designing and establishing a multi-view camera studio tailored for dataset capture
- Defining dataset details, participant guidelines, and collection protocols
- Developing custom error filtering and sampling toolkits for data refinement
- Implementing an optimization framework to extract accurate 3D poses from RGB-Depth data
- Generating precise ground-truth via iterative mesh-based rendering workflows including hybrid verification step
Quality Control & Validation
To maximize data quality:
- Utilized a bootstrapping approach to reject the outlier pseudo GT keypoints
- Fine-tuned the segmentation model per object with our manually annotated segmentation masks
- Conduct both automatic and manual verification steps to further filter out the noisy annotations
Role & Impact
As the general lead, I navigated both technical and organizational challenges—solving engineering problems while driving efficient communication between companies. The project was successfully completed within a short timeline thanks to cohesive teamwork and agile execution. This hands-on experience in large-scale AI data construction and multi-stakeholder coordination has become a cornerstone of my skillset for future endeavors.
Cite
---
@inproceedings{cho2024dense,
title={Dense hand-object (ho) graspnet with full grasping taxonomy and dynamics},
author={Cho, Woojin and Lee, Jihyun and Yi, Minjae and Kim, Minje and Woo, Taeyun and Kim, Donghwan and Ha, Taewook and Lee, Hyokeun and Ryu, Je-Hwan and Woo, Woontack and others},
booktitle={European Conference on Computer Vision},
pages={284--303},
year={2024},
organization={Springer}
}
---