Suekyeong
Nam
About Me
CV | Blog | akasha9890@gmail.com
I received my MS and BS degrees in Computer Science and Engineering department at Kyung Hee University in South Korea advised by Seungkyu Lee. My research focuses on 3D human pose estimation and capturing 3D human pose data at a motion capture studio.
Publications
JT-MGCN: Joint-temporal Motion Graph Convolutional Network for Skeleton-Based Action Recognition
Suekyeong Nam, Suengkyu Lee
IEEE International Conference on Pattern Recognition (ICPR) 2020
PDF | Slides | Poster
Suekyeong Nam, Suengkyu Lee
IEEE International Conference on Pattern Recognition (ICPR) 2020
PDF | Slides | Poster
Motion regeneration using motion texture and autoencoder
Suekyeong Nam, Seungkyu Lee
SIGGRAPH Asia 2018 (Poster Session)
PDF
Suekyeong Nam, Seungkyu Lee
SIGGRAPH Asia 2018 (Poster Session)
On-line Color Model Learning for Adaptive Skin Segmentation.
Suekyeong Nam, Seungkyu Lee
Korea Computer Congress 2015
PDF
Suekyeong Nam, Seungkyu Lee
Korea Computer Congress 2015
Human Body Related Research Experiences
Multi-view Multi-persons 3D Pose Estimation
- Predict human poses using multiview images.
- Predict human poses using multiview images.
3D Personal Avatar Generation for Augmented Reality
- Predict human poses in real-time using voxel datasets, which are extracted from the multi-view depth camera images. - Infer 3D personal body shapes and textures based on 2D multi-view images.
- Predict human poses in real-time using voxel datasets, which are extracted from the multi-view depth camera images. - Infer 3D personal body shapes and textures based on 2D multi-view images.
Action Recognition Methods Effective for Both Spatial and Temporal Features
- Started with extracting spatial and temporal features related to symmetrical movement only first.This method manually calculated the features along the spatial-temporal axis from 2D motion textures that are transformed from 3D skeleton data. - The follow-up work used CNN and LSTM to extract a variety of both spatial and temporal features. CNN Autoencoder reduced noise and dimensions on 2D motion textures. The compressed texture information enabled to utilize LSTM on images. Results published in 2018 SIGGRAPH Asia poster session. - The last work used an adaptive graph structure instead of 2D motion textures, based on my observation of strong correlation between spatial and temporal features. Unlike prior work, this work generated inter-frame edges adaptively. Results will be published in ICPR 2020.
- Started with extracting spatial and temporal features related to symmetrical movement only first.This method manually calculated the features along the spatial-temporal axis from 2D motion textures that are transformed from 3D skeleton data. - The follow-up work used CNN and LSTM to extract a variety of both spatial and temporal features. CNN Autoencoder reduced noise and dimensions on 2D motion textures. The compressed texture information enabled to utilize LSTM on images. Results published in 2018 SIGGRAPH Asia poster session. - The last work used an adaptive graph structure instead of 2D motion textures, based on my observation of strong correlation between spatial and temporal features. Unlike prior work, this work generated inter-frame edges adaptively. Results will be published in ICPR 2020.
Evaluation of Gesture Recognition Algorithms for Embedded Devices
- Compared, on an actual embedded AI computing device, the CNN-based algorithm that the project developed for 2D images against basic computer vision and other CNN-based algorithms.
- Compared, on an actual embedded AI computing device, the CNN-based algorithm that the project developed for 2D images against basic computer vision and other CNN-based algorithms.
Research Experiences
Automatic Color Correction on 2D Product Images for Ecommerce
Identified corresponding points across two different image dimensions, such as plain product shots and lifestyle images. Then corrected the colors on the lifestyle images to the ones in the other using transformation networks. Results
Identified corresponding points across two different image dimensions, such as plain product shots and lifestyle images. Then corrected the colors on the lifestyle images to the ones in the other using transformation networks. Results
Real-time Color Model Learning for Adaptive Skin Segmentation.
Segmented personal skin using adaptive Gaussian distribution model for colors on videos. PDF
Segmented personal skin using adaptive Gaussian distribution model for colors on videos. PDF
Honors and Awards
Mosaicer: Automatic Pixelation of Untargeted Faces on Videos
Suekyeong Nam, Seongah Jo
Grand Prize, NAVER crop.
Code | Demo video | Slides
Suekyeong Nam, Seongah Jo
Grand Prize, NAVER crop.
Code | Demo video | Slides
Entrance Examination Scholarship(Merit-based)
Kyung Hee University, 02/2018–08/2019
Kyung Hee University, 02/2018–08/2019
Excellence Award (Graduated with the highest honor in my class)
Kyung Hee University, 02/2018
Kyung Hee University, 02/2018
National Science Technology scholarship
Korea Student Aid Foundation, 03/2016–08/2017
Korea Student Aid Foundation, 03/2016–08/2017
Superiority Scholarship(Merit-based)
Kyung Hee University, 08/2014–02/2015
Kyung Hee University, 08/2014–02/2015