Sung-Un Park

I am Sung-Un Park, a staff engineer at Samsung AI Center (Process/Equipment AI Lab) and a Ph.D. student in Electrical Engineering at KAIST (Computational Intelligence Lab, advised by Prof. Changick Kim). My work focuses on building reliable, deployable AI systems in computer vision and on-device ML. At Samsung, I contributed to several Galaxy features, including face recognition and Intelligent Scan(multimodal face recognition), fingerprint anti-spoofing, and the Bixby wake-up/unlock speaker-recognition pipeline. More recently, I have worked on semiconductor manufacturing use cases, developing 3D depth metrology from SEM images and machine-learning–based virtual metrology. I also maintain an interest in robust learning for noisy and imbalanced data, connecting research with product needs.

selected publications

ICML

PRIME: Deep Imbalanced Regression with Proxies

Jongin Lim, Sucheol Lee, Daeho Um, and 2 more authors

In Forty-second International Conference on Machine Learning, 2025

HTML PDF
Conference

A Depth Camera-based Human Activity Recognition via Deep Learning Recurrent Neural Network for Health and Social Care Services

S.U. Park, J.H. Park, M.A. Al-masni, and 3 more authors

Procedia Computer Science, 2016

International Conference on HCist 2016

Abs DOI

Human activity recognition (HAR) has become an active research topic in the fields of health and social care, since this technology offers automatic monitoring and understanding of activities of patients or residents. Depth camera-based HAR recognizes human activities using features from depth human silhouettes via conventional classifiers such as Hidden Markov Model (HMM), Conditional Random Fields etc. In this paper, we propose a new HAR system via Recurrent Neural Network (RNN) which is one of deep learning algorithms. We utilize joint angles from multiple body joints changing in time which are represented a spatiotemporal feature matrix (i.e., multiple body joint angles in time). With these derived features, we train and test our RNN for HAR. In order to evaluate our system, we have compared the performance of our RNN-based HAR against the conventional HMM- and Deep Belief Network (DBN)-based HAR using a database of Microsoft Research Cambridge-12 (MSRC-12). Our test results show that the proposed RNN-based HAR is able to recognize twelve human activities reliably and outperforms the HMM- and DBN-based HAR. We have achieved the average recognition accuracy of 99.55% for the activities. The results are 7.06% more accurate than that of the HMM-based HAR and 2.01% more accurate than that of the DBN-based HAR.