CV

Hi! I'm Seulgi Kim

My research aims to explain Why AI understands video. Specifically, I aim to advance computer vision and multi-modal learning algorithms to interpret temporal and semantic relations between actions in long-term videos.

I am a second year Ph.D. student at Georgia Tech, advised by Ghassan AlRegib. I received my B.S. from Seoul National University, where I was supported by Hyundai Motor Group Research Scholarship.

Before joining Tech, I worked as a Research Engineer at Hyundai Motor Company for 3 years, where I developed AI models for perception and localization in autonomous driving. We launched Lv.4 Autonomous Driving Pilot Services for RoboTaxi, RoboShuttle, and RoboTruck.

Recent News

Featured Research Publications

Multi-level and Multi-modal Action Anticipation

Action anticipation framework using multi-modal and hierarchical modeling

ICIP 2025

Hierarchical and Multimodal Data for Daily Activity Understanding

Multimodal, multi-view, multiple levels of Human Activity Dataset collaborating with Amazon

DMLR 2025

HEX: Hierarchical Emergence Exploitation in Self-Supervised Algorithms

Self-supervised (SSL) algorithm to take advantage of hierarchical structures that emerge during training

WACV 2025

Education

Work Experience

• Developed a Generative AI model that auto-tag the root causes of 1000+ seller inquiries.
• Built a chatbot to be used by 100+ AWS Partner Organization (APO) developers using Claude 3 Sonnet LLM.
• Fixed the team's PSF pipelines; identified the root causes that cause failure and rolled out to production stages.

• Developed a real-world 3D representation model for detecting roads and road markers using monocular camera and 3D geometric priors.
• Responsible for the entire perception pipeline for the Level 4 self-driving pilot service, including dataset creation, neural network implementation, post-processing with 3D camera geometry, and evaluation. Launched on Hyundai Robo-Taxi, Robo-Shuttle, and Robo-Trucks services.
• Published 12 Patents and an in-house conference paper (Excellence Paper (Top 10)) by fusing the LiDAR and Camera modalities.
• Built a debugging monitor from scratch using QT5 to communicate with embedded systems via TCP/IP.
• Processed a raw data acquired from camera, LiDAR, and CAN bus sensors to train AI models.

• Developed a Drivable Surface Segmentation and judgment model using CAN signal and monocular camera images.

Publications

			
@article{kim2025multi,
  title={Multi-level and Multi-modal Action Anticipation},
  author={Kim, Seulgi and Kaviani, Ghazal and Prabhushankar, Mohit and AlRegib, Ghassan},
  journal={arXiv preprint arXiv:2506.02382},
  year={2025}
}
		

			
@article{kaviani2025hierarchical, 
  title={Hierarchical and Multimodal Data for Daily Activity Understanding}, 
  author={Kaviani, Ghazal and Yarici, Yavuz and Kim, Seulgi and Prabhushankar, Mohit and AlRegib, Ghassan and Solh, Mashhour and Patil, Ameya}, journal={arXiv preprint arXiv:2504.17696}, 
  year={2025} 
}
		

			@article{WACV2025,
  author  = {Kiran Kokilepersaud and Seulgi Kim and Mohit Prabhushankar and Ghassan AlRegib},
  title   = {HEX: Hierarchical Emergence Exploitation in Self-Supervised Algorithms},
  journal = {Winter Applications of Computer Vision (WACV)},
  year    = {2025},
  url     = {https://arxiv.org/pdf/2410.23200?}
}

		

Teaching

Honors and Awards

Patents

Service

Skills

Programming: Python, C++, C, Matlab, R, Java, Ocaml

Deep Learning Frameworks: PyTorch, TensorFlow, Keras, TensorRT, Cuda Programming

Tools: Open3D, CAN, TCP/IP, OpenMP, OpenGL, AWS, Git, JIRA, Docker, Raspberry Pi, Arduino, Android Studio, Solidworks

Web and Writing: HTML, PHP, CSS, JavaScript, LaTeX, Markdown, Git

OS: Mac OS X, Ubuntu, Unix Command Line, Windows

Book

References