Research Project
Monocular 3D Object Detection
1. RARE: Learning to RAnk and REtrieve for Monocular 3D Object Detection
-
CVPR 2026 (Highlight)
-
RARE rethinks confidence learning in monocular 3D detection by casting it as a ranking problem rather than score regression. By aligning predictions with relative geometric quality, we obtain more stable and reliable confidence estimates. We further model multimodal 3D ambiguity through diverse hypothesis generation and retrieval.
(a) Conventional method
(b) ours
Overall architecture of RARE
2. Modeling and Learning Multiple Hypotheses for Monocular 3D Object Detection
-
WACV 2026
-
Proposed multi-hypothesis generation, learning, and filtering.
-
generate multiple hypotheses for uncertain objects to handle ill-posedness of the task
-
Ear Recognition
-
Developed a fully automated ear recognition system based on deep learning
- full pipeline: data acquisition, ear detection, feature
extraction, and matching (identification/verification)
- self-attention and margin-based angular loss
-
Introduced in YTN (news)
- https://www.youtube.com/watch?v=yAbYLXd5acs&t=89s -
Introduced in MBC (news)
- (05:30) https://www.youtube.com/watch?v=-p3pD1eEF9Q&t=300s
-
Developed a fully automated ear recognition system based on deep learning
- full pipeline: data acquisition, ear detection, feature
extraction, and matching (identification/verification)
- self-attention and margin-based angular loss
-
Introduced in YTN (news)
- https://www.youtube.com/watch?v=yAbYLXd5acs&t=89s -
Introduced in MBC (news)
- (05:30) https://www.youtube.com/watch?v=-p3pD1eEF9Q&t=300s
Video Anomaly Detection
-
Developed an algorithm for improving the accuracy of violence detection in public transportation systems
-
Developing an unsupervised/weakly-supervised video anomaly detection algorithm
- delving into hard sample mining algorithms for
contrastive learning to enhance the anomaly
localization performance in untrimmed videos
-
Developed an anomaly detection program
- used model (PatchCore, CVPR 2022) provided here -
Developing a defect detection algorithm with the industrial image datasets
Defect Detection
Pet Face Recognition
-
Developed a fully automated deep learning-based pet face recognition system
-
Applied self-attention mechanism to get a model robust to occlusion
Food Recognition System
-
Developed a real-time food recognition system
-
Applied computer vision techniques such as blob detection, erosion, and dilation for plate detection
KIST-Face Database Construction
-
Participated in constructing a large-scale Korean face database (KFACE database) at KIST
-
Data acquisition, and post-processing such as data cleansing and annotating
KIST-Structured Ear Database Construction
-
Constructed a large-scale ear database based on the KFACE dataset