Aro Kim
Generative Models / Video Coding / 3D Vision
Fourth-year Integrated M.S./Ph.D. student in Computer Science and Engineering at Kyungpook National University, advised by Prof. Sang-hyo Park.
About
My research focuses on generative models, video coding, and 3D vision. I am particularly interested in developing efficient and high-performance methods for real-world image and video understanding.
Research Approach
- I enjoy quickly learning new research ideas and codebases, adapting them to my own work, and iteratively improving performance.
- I value open communication and collaborative problem-solving, and enjoy working closely with colleagues throughout the research process.
Education
Integrated M.S./Ph.D. Student in Computer Science and Engineering
Fourth-year student in the Integrated M.S./Ph.D. Program
Advisor: Prof. Sang-hyo Park
Research Interests
- Generative Models
- Image and Video Restoration
- Video Coding
- 3D Vision
Publications
FiDeSR: High-Fidelity and Detail-Preserving One-Step Diffusion Super-Resolution
CVPR 2026
A one-step diffusion super-resolution method designed to preserve high-fidelity details while maintaining efficient inference.
CompSplat: Compression-aware 3D Gaussian Splatting for Real-world Video
BMVC 2026
A compression-aware 3D Gaussian Splatting framework for long, unposed, real-world videos. It models frame-wise compression reliability and adaptively guides Gaussian densification and pruning to improve rendering quality, pose accuracy, and geometric consistency under severe compression.
Deep Learning-Guided Video Compression for Machine Vision Tasks
EURASIP Journal on Image and Video Processing, 2024, Vol. 2024, No. 1, Article 32
This work proposes a video compression framework tailored to machine vision tasks. It applies encoders to video regions distinguished by machine vision to improve coding efficiency, achieving an average BD-rate gain of 5.91% and up to 19.51% BD-rate gain.
Pruning-Guided Feature Distillation for an Efficient Transformer-Based Pose Estimation Model
IET Computer Vision, 2024, Vol. 18, No. 6, pp. 745–758
This work proposes a pruning-guided feature distillation strategy for an efficient transformer-based 3D human pose estimation model. The approach reduces model size by 30% compared to the state of the art while maintaining high accuracy.
Patents
Apparatus for Generating High Quality Image and Method Thereof
KR Patent No. 10-2838222 · Granted July 21, 2025
A training framework for high-quality image generation that adaptively adjusts channel configurations based on output quality during training.
Contact
Email: arokim37@gmail.com / arokim37@knu.ac.kr