KyoungJin Oh

I am an undergraduate student in Computer Science and Engineering at Yonsei University. Currently, I am a research intern at the KAIST Computer Vision Lab (CVLAB), where I contribute to the image and video restoration team.
I aspire to become an impactful researcher by producing work that is broadly applicable across various sub-fields of Computer Vision.

Email  /  CV  /  Github

Research Interests

I have a broad interest in various fields of Computer Vision. Currently, as a member of the restoration team, I have been conducting research in this area. Recently, I studied Text-Aware Image Restoration (TAIR), a new task focused on restoring both visual quality and text fidelity in degraded images. Building on this, I am currently studying video restoration using optical flow. In the future, I plan to delve into diffusion guidance.

Research Experience

• Advisor: Prof. Seungryong Kim
• Image & Video Restoration

• Advisor: Prof. Seonjoo Kim
• Diffusion models

Education

• B.S. of Computer Science and Engineering
  Mar 2022 - Present

Honors & Awards

High Honors: Top 3%; Honors: Top 10% of the cohort

• Honors
  Awarded by Yonsei university, 2nd semester of 2023


• Silver Prize, The 1st Yonsei GenAI Contest
  KyoungJin Oh, JunSeong Lee, KangWon Lee
  Awarded by Yonsei University SW-Centered University Program, Dec 2023

  We developed web-based software using the OpenAI API. The software takes an artwork as input and outputs the appropriate lighting for the exhibition and how the artwork would appear when displayed in the exhibition under that lighting.



• High Honors
  Awarded by Yonsei university, 1st semester of 2023


• Honors
  Awarded by Yonsei university, 1st semester of 2022

Course Projects

I implemented image filters in both the spatial and frequency domains to restore noisy images. I manually built spatial filters like Gaussian and Laplacian via convolution, and also implemented frequency filters using Fourier transforms, ultimately synthesizing these techniques for practical image denoising.

I used PCA with SVD to extract Eigenfaces (principal components) for Face Recognition. The required number of components was determined by a cumulative variance threshold. Face images were reconstructed by projecting them onto the PCA basis. Final classification utilized the nearest neighbor algorithm to match test images in the PCA space via Euclidean distance.

Homography matrix H was estimated from correspondence points by applying SVD to Matrix A. RANSAC was used for robust estimation to maximize inliers (points within an error threshold). Using more correspondences increased the method's robustness. The main limitation is the violation of the planarity assumption in 3D scenes, which causes artifacts due to parallax.

This project proposed a method for Sentiment Classification of Korean sentences by sequentially using the NLLB-200-3.3B (NLLB) model to translate Korean into English and the English sentiment analysis model, twitter-roberta-base-sentiment-latest (ROBERTa). The proposed model achieved an Accuracy of 0.7859, demonstrating better performance than the comparison model, bert-base-cased-Korean-sentiment, which scored 0.7247.

Extracurricular Activities

As the Vice President of Uni-D, I planned and executed key aspects of Uni-DTHON—a hackathon and datathon competition organized by Uni-D, an organization formed by the student councils of various university computer science departments—which included theme development, data curation, and venue selection.

As the President of the Student Council for the Department of Computer Science and Engineering, I managed and administered various department events. Furthermore, I effectively collaborated with diverse individuals, building effective relationships throughout this role.

I competed in events such as K7, SUFA, and the Yonsei University President’s Cup. Through participating in multiple matches, I developed teamwork skills.

I learned how to generate and develop business ideas in the entrepreneurial space through projects like international startup benchmarking, and contributed to a project aimed at developing a new 3D modeling tool program as an Minimum Viable Product (MVP).