Experiences
2022.05 ~ 2022.11 within M.S.
(Research Project)
Advisor: Yukyung Choi
Development of MPEG CDVA-based Video Search Software (Ongoing)
Related Areas: Video Retrieval (CBVR)
Co-worker: Guentaek Lim, Hyunwoo Kim, (Sejong RCV), Joonsoo Kim, Hyongon Choo (ETRI)
Objective: Development of efficient video learning techniques for untrimmed videos.
2021.04 ~ 2021.11 within M.S.
(Research Project)
Advisor: Yukyung Choi
Dark Data Analysis Technology for Data Scale and Accuracy Improvement
Related Areas: Active Learning, Self-Learning
Co-worker: Yujin Hwang, Guentaek Lim (Sejong RCV)
Objective: Development of data value judgment technology for dark data analysis.
The whole process is made up of two parts: 1) an investigation of active learning, which chooses the order of learning data; and 2) an investigation of self-learning, which learns with the label made by the model.
My main role is to solve the gradient conflict problem that arises when dealing with multi-tasks in active learning through an investigation of gradient surgery.
2021.04 ~ 2021.11 within M.S.
(Research Project)
Advisor: Yukyung Choi
A Study on a Deep Learning Model that Extracts Video Visual Features for Machine Vision Task
Related Areas: Video Retrieval (CBVR), Video Alignment
Co-worker: Yujin Hwang, Guentaek Lim (Sejong RCV), Joonsoo Kim, Joungil Yun (ETRI)
Objective: Development of an efficient video retrieval technique for untrimmed videos.
The whole process is made up of two parts: 1) the creation of a foreground selection module to eliminate unnecessary video areas; and 2) the investigation of a framework for simultaneous video retrieval and alignment.
My main role encompasses the entire process.
2020.04 ~ 2020.11 within B.S.
(Research Project)
Advisor: Yukyung Choi
A Study on a Deep Learning Model for Extracting Visual Features by Machine Vision Task
Related Areas: Image Retrieval, Video Retrieval (CBVR)
Co-worker: Daechan Han, Chanho Jeong (Sejong RCV), Joonsoo Kim, Joungil Yun (ETRI)
Objective: CDVA framework analysis and enhancement for video retrieval.
The whole process is made up of two parts: 1) analysis of how the CDVA framework describes and compresses features; and 2) the development of a better descriptor for video retrieval.
My main role encompasses the entire process.
Related Material
Exploring the Temporal Cues to Enhance Video Retrieval on Standardized CDVA
Compression Method for MPEG CDVA Global Feature Descriptors
Joonsoo Kim, Won Jo, Guentaek Lim, Joungil Yun, Sangwoon Kwak, Soon-heung Jung, Won-Sik Cheong, Hyon-Gon Choo, Jeongil Seo, and Yukyung Choi
Journal of Broadcast Engineering
[Paper]
2020.04 ~ 2020.06 within B.S.
(Domestic Challenge)
Advisor: Yukyung Choi
NAVER LABS Mapping & Localization Challenge 2020: Indoor Track
3rd place (Participants: 158, Participating teams: 115)
Related Areas: Visual Localization (Image Retrieval, Pose Estimation)
Co-worker: Daechan Han, Taejoo Kim, Jiwon Kim, Yujin Hwang (Sejong RCV)
Objective: Estimation of camera position and orientation (6DoF) for query images in an indoor environment.
The whole process is made up of four steps: 1) retrieval of images; 2) reranking; 3) local matching; and 4) pose estimation
My main role lies in image retrieval, reranking, and pose estimation.
2018.11 ~ 2020.04 within B.S.
(Research Project)
Advisor: Yukyung Choi
Visualization of Thermal Image and Development of Perception Sensors based on Deep Learning
Related Areas: Sensor Fusion, Multispectral Object Detection
Co-worker: Jeongmin Shin, Jiwon Kim, Daechan Han, Taejoo Kim, Yujin Hwang, Hyunho Nam, Byungjoo Kim, Namhoon Kim, Chanho Jeong, Jinsoo Lee (Sejong RCV)
Objective: Establishment of a system to improve the safety of unmanned forklifts through sensor fusion.
The whole process is made up of four steps: 1) configuration, synchronization, and calibration of multiple sensors (RGB-Thermal-RGB-Thermal); 2) acquisition a multi-sensor dataset with depth estimation and alignment 3) model lightweight via AMP; 4) demonstration test in a logistics warehouse.
My main role lies in obtaining a dataset through stereo calibration, depth estimation, and image alignment among multi-sensor cameras.
Related Material
R2T2: RGB-Thermal-Depth Dataset for Pedestrian Detection
Taejoo Kim*, Jeongmin Shin*, Won Jo*, Daechan Han*, Jiwon Kim, Byungjoo Kim, Hyunho Nam, Yujin Hwang, Namhoon Kim, and Yukyung Choi
32nd Workshop on Image Processing and Image Understanding (IPIU)
[Paper], [Code](RGB-Thermal Alignment Code)
2018.09 ~ 2020.08 within B.S.
(Research Project)
Advisor: Yukyung Choi
A Fault-tolerant Fusion Method for Robust Pedestrian Detection
Related Area: Multispectral Object Detection
Co-worker: Jiwon Kim, Daechan Han, Hyunho Nam, Jeongmin Shin, Taejoo Kim, Yujin Hwang, Byungjoo Kim (Sejong RCV)
Objective: Development of a pedestrian detection model that is robust in normal and abnormal conditions.
The whole process is made up of two parts: 1) design of a pedestrian detection model with sensor fusion (RGB-Thermal and RGB-Thermal-Depth) in normal conditions; 2) design of that model in abnormal conditions.
My main role lies in designing a pedestrian detection model with sensor fusion in normal and abnormal conditions.
Related Material
Single-Shot Adaptive Fusion Network for Robust Multispectral Pedestrian Detection
Jiwon Kim*, Won Jo*, Hyunho Nam, Soonmin Hwang, Chiwon Roh, Namil Kim, and Yukyung Choi
32nd Workshop on Image Processing and Image Understanding (IPIU), Best Paper Award
[Paper]
2019.04 ~ 2019.06 within B.S.
(In Class)
Advisor: Yukyung Choi
Sign Language Classification
Related Area: Action Recognition
Co-worker: Daechan Han, Miyeon Lee, Joonhui Park, Kyuwon Lee (Sejong Univ.)
Objective: Development of a sign language classifier for communication with the deaf.
The whole process is made up of three steps: 1) sign language dataset acquisition; 2) hand position detection; and 3) classifier design.
My main role lies in classifier design and the acquisition of a sign language dataset.