Taeyoung Son

Seoul, South Korea

I completed my Master’s Degree at Computer Vision Laboratory at the Dept. of Computer Science and Engineering of the POSTECH, advised by Prof. Suha Kwak. I also did my BS at CSE of POSTECH.

I am a machine learning engineer specializing in bridging the gap between state-of-the-art research and scalable, production-ready products. My work spans recommendation systems, generative AI, and computer vision, combining strong research experience with hands-on engineering expertise. My interests include artificial intelligence, machine learning, computer vision, and software design and development. My research experiences encompass generative AI, domain adaptation and generalization, 3D human mesh reconstruction, and image recognition under extreme conditions (e.g., rain, frost, snow, etc.). I am currently a machine learning engineer at RIDI.

News

Oct 23, 2023	🏢 I joined RIDI as a machine learning engineer.
Feb 1, 2023	📝 A paper on domain generalization for semantic segmentation is accepted to ICRA 2023.
Nov 7, 2022	🥳 I won the Qualcomm Innovaation Fellowship 2022.
Jun 11, 2022	📝 A paper on foggy scene semantic segmentation is accepted to CVPR.
Apr 4, 2022	🏢 I joined NALBI as a research scientist.
Nov 5, 2020	📝 A paper on visual recognition under extreme condition is accepted to ECCV.

Education

Mar, 2020 - Feb, 2022	Pohang University of Science and Technology (POSTECH), Pohang, South Korea M.S. in Computer Science and Engineering Advisor: Prof. Suha Kwak
Mar, 2015 - Feb, 2020	Pohang University of Science and Technology (POSTECH), Pohang, South Korea B.S. in Computer Science and Engineering

Experience

October, 2023 - Present	RIDI, Seoul, South Korea Machine Learning Engineer Implemented pipelines for ControlNet, LoRA, and ViTPose models. Developed a inpaint model for Prodifi. Developed a recommendation system for the ‘For You’ section of MANTA. Developed a recommendation system for the ‘Trending Now’ section of RIDIBOOKS. Developed RAG chatbot ‘GONYAGI’, a real-time translation chatbot integrated with Slack.
April, 2022 - October, 2023	NALBI, Seoul, South Korea AI Research Scientist Developed a real-time 3D human body reconstruction algorithm. Implemented the Part Attention Regressor for 3D Human Body Estimation (PARE) model. Automated SMPL annotation pipeline using ROKOKO motion capture equipment. Optimized models for real-time inference on mobile and CPU-only environments.
June, 2019 - Aug, 2019	HyperConnect, Seoul, South Korea Research Scientist Internship Trained a classification model to filter inappropriate profile pictures in Azar. Developed an image classification model to automatically filter inappropriate images from a large-scale, imbalanced dataset over millions of images. Improved the existing Inception-ResNet V2 model by replacing it with EfficientNet-V5, achieving an accuracy of 96.51%. Researched and applied various self-supervised learning techniques, such as Data Distillation and Label Refinery, to address data imbalance and noise issues.
June, 2018 - Sep, 2018	HyperConnect, Seoul, South Korea Research Scientist Internship Developed a model to generate Look-Up Tables (LUTs) for on-device camera beautification filters. Implemented a web crawler to collect highly-rated images from Flickr. Developed a pixel-wise LUT generator for image enhancement and color correction.

Publications

WEDGE: Web-Image Assisted Domain Generalization for Semantic Segmentation

Namyup Kim, Taeyoung Son, Jaehyun Pahk, Cuiling Lan, Wenjun Zeng, and Suha Kwak

IEEE International Conference on Robotics and Automation (ICRA), 2023

Abstract Paper Website

Domain generalization for semantic segmentation is highly demanded in real applications, where a trained model is expected to work well in previously unseen domains. One challenge lies in the lack of data which could cover the diverse distributions of the possible unseen domains for training. In this paper, we propose a WEb-image assisted Domain GEneralization (WEDGE) scheme, which is the first to exploit the diversity of web-crawled images for generalizable semantic segmentation. To explore and exploit the real-world data distributions, we collect a web-crawled dataset which presents large diversity in terms of weather conditions, sites, lighting, camera styles, etc. We also present a method which injects the style representation of the web-crawled data into the source domain on-the-fly during training, which enables the network to experience images of diverse styles with reliable labels for effective training. Moreover, we use the web-crawled dataset with predicted pseudo labels for training to further enhance the capability of the network. Extensive experiments demonstrate that our method clearly outperforms existing domain generalization techniques.
FIFO: Learning Fog-Invariant Features for Foggy Scene Segmentation

Sohyun Lee, Taeyoung Son, and Suha Kwak

IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

Abstract Paper Code Website

Robust visual recognition under adverse weather conditions is of great importance in real-world applications. In this context, we propose a new method for learning semantic segmentation models robust against fog. Its key idea is to consider the fog condition of an image as its style and close the gap between images with different fog conditions in neural style spaces of a segmentation model. In particular, since the neural style of an image is in general affected by other factors as well as fog, we introduce a fog-pass filter module that learns to extract a fog-relevant factor from the style. Optimizing the fog-pass filter and the segmentation model alternately gradually closes the style gap between different fog conditions and allows to learn fog-invariant features in consequence. Our method substantially outperforms previous work on three real foggy image datasets. Moreover, it improves performance on both foggy and clear weather images, while existing methods often degrade performance on clear scenes.
Urie: Universal image enhancement for visual recognition in the wild

Taeyoung Son, Juwon Kang, Namyup Kim, Sunghyun Cho, and Suha Kwak

European Conference on Computer Vision (ECCV), 2020

Abstract Paper Code Website

Despite the great advances in visual recognition, it has been witnessed that recognition models trained on clean images of common datasets are not robust against distorted images in the real world. To tackle this issue, we present a Universal and Recognition-friendly Image Enhancement network, dubbed URIE, which is attached in front of existing recognition models and enhances distorted input to improve their performance without retraining them. URIE is universal in that it aims to handle various factors of image degradation and to be incorporated with any arbitrary recognition models. Also, it is recognition-friendly since it is optimized to improve the robustness of following recognition models, instead of perceptual quality of output image. Our experiments demonstrate that URIE can handle various and latent image distortions and improve the performance of existing models for five diverse recognition tasks when input images are degraded.

Honors and Awards

Qualcomm Innovation Fellowship South Korea (2022)

Winner ($3,000) - FIFO: Learning fog-invariant features for foggy scene segmentation (CVPR2022)

CVPR Best Paper Finalist (2022)

Our paper on foggy scene segmentation is nominated as a best paper finalist in CVPR 2022.
Awarded to Top 0.4% (33 of 8161 papers)
FIFO: Learning Fog-invariant Features for Foggy Scene Segmentation

Seoul MaaS Hackerthon (2019)

Won excellence award at MaaS hackerthon hosted by the city of seoul