Byeongju Woo

I am a first lieutenant in the Republic of Korea Air Force (ROKAF), currently serving as a research officer at the Agency for Defense Development (ADD), the Korean counterpart to the U.S. DARPA. Previously, I completed my bachelor's degree (Summa Cum Laude) in Computer Science and Engineering at POSTECH.

My research interests lie in computer vision and deep learning. I've worked on the robust visual recognition in adverse conditions and Out-of-Distribution generalization.

Email  |  CV  |  Scholar  |  LinkedIn

profile photo

Research Interests

My research goal is to improve reliability of AI systems by establishing confidence in their behavior and robustness. To this end, I am currently focusing on the below topics, but not limited to:

  • Robust Visual Recognition
  • Out-of-Distribution (OOD) Generalization
  • Uncertainty Estimation
  • Diagnosing and Rectifying Deep Learning model

Publications
Descriptive Alt Text
Textual Query-Driven Mask Transformer for Domain Generalized Segmentation
Byeongju Woo*, Byeonghyun Pak*, Sunghwan Kim*, Daehwan Kim, Hosung Kim
ECCV, 2024
Project page | arXiv

We introduce a semantic segmentation method that generalizes well to unseen domains. Our method effectively handles domain shifts by exploiting domain-invariant semantics from text embeddings of vision-language models (e.g., CLIP).

Descriptive Alt Text
Human Pose Estimation in Extremely Low-Light Conditions
Sohyun Lee*, Jaesung Rim*, Boseung Jeong, Geonu Kim, ByungJu Woo, Haechan Lee, Sunghyun Cho, Suha Kwak
CVPR, 2023
Project page | arXiv | Code

We study human pose estimation in extremely low-light images, where severely corrupted inputs degrade prediction significantly. We build a dataset that provides real and aligned low-light and well-lit images and present a strong baseline method that fully exploit them.

Projects
Descriptive Alt Text
Adversarial Patches with Enhanced Camouflage against Soldier Detection
Ministry of Science and ICT, 2021

We develop adversarial patches, to hide from the deep learning models that detect allies. We also aim for patches to resemble military uniform patterns, in order to camouflage well to humans.

Descriptive Alt Text
Object Detection for 3D NAND Flash Failure Detection
SK Hynix, department of DS Algorithm, 2021

We develop an automated algorithm to pre-check for rejects in 3D NAND Flash manufacturing. To this end, we design an object detection model that is reliable in the industrial field, with a small number of training data.

Award and Scholarship

  • Excellence Research Award (2022), POSTECH
    • Selected as the 2nd place among 72 undergraduate CSE research projects
  • Minister Award (2021), Ministry of Science and ICT
    • Selected as the 1st place among 6 defense research projects

website template from Jon Barron