Inseop Chung

I'm a Staff Researcher at the AI Center of Samsung Electronics, where I work on various topics in computer vision and machine learning.

I received my PhD from Graduate School of Convergence Science and Technology (GSCST), Department of Intelligence and Information of Seoul National University (SNU) under the supervision of professor Nojun Kwak. I was a member of the Machine Intelligence and Pattern Analysis Lab (MIPAL) led by Prof. Kwak. My research focuses on domain adaptation, domain generalization, test-time adaptation, object detection, semantic segmentation, knowledge distillation, and vision-language-action models.

Email  /  CV  /  Google Scholar  /  LinkedIn  /  GitHub

Research Interests

I'm broadly interested in computer vision, deep learning, and machine learning, with a current and primary focus on Vision-Language-Action (VLA) models for embodied AI and robotics. My goal is to build AI systems that can perceive, reason, and act robustly in dynamic, real-world environments.

My recent research centers on making AI models more adaptive and generalizable to unseen scenarios through techniques such as domain adaptation, domain generalization, and continual test-time adaptation (CTA). I also have experience in knowledge distillation, object detection, and semantic segmentation.

Currently, I'm particularly passionate about advancing VLA models that integrate vision and language with action understanding for real-world robotic control. I actively work with open-source VLA frameworks such as OpenVLA and Physical Intelligence (OpenPI) models, and have hands-on experience in adapting and extending these platforms for robotic manipulation tasks.

My long-term interest lies in developing scalable and generalizable embodied intelligence — enabling agents that can learn continuously from diverse environments, understand natural language instructions, and perform complex physical tasks with minimal supervision.

Publications

This work improves Continual Test-Time Adaptation by introducing class-wise feature clustering and source-aligned prototype anchoring. It dynamically reduces prediction bias and enhances model robustness under distribution shifts.

We propose a single-network method for Open Domain Generalization that improves unseen domain generalization and unknown class detection. It leverages a linear-probed head and two regularization terms to refine the classifier while preserving pre-trained features, enabling efficient and robust generalization.

We propose a lightweight, architecture-agnostic test-time adaptation method for object detection under continual domain shifts. By updating only small adaptor modules and introducing a class-wise feature alignment strategy, our approach achieves efficient, robust adaptation without modifying the backbone, improving mAP on standard benchmarks while preserving real-time performance.

We introduce D-RMM, an end-to-end multi-object detector trained with negative log-likelihood and a novel regularization loss. By modeling detection as density estimation, our method removes training heuristics, improves confidence reliability, and outperforms prior end-to-end detectors on MS COCO.

We enhance UDA for semantic segmentation by transferring inter-pixel correlations from the source to the target domain using a self-attention module. Trained only on the source, this module guides the segmentation network on the target domain beyond noisy pseudo labels, improving performance on standard benchmarks.

We introduce X-MAS, a large-scale multi-modal dataset for outdoor surveillance with 500K+ annotated image pairs across five sensor types.

We propose OADA, an offset-aware domain adaptive object detector that conditions feature alignment on bounding box offsets to reduce negative transfer. This simple yet effective method improves both discriminability and transferability, achieving state-of-the-art results in domain adaptive detection.

We propose D-ProtoNets, a metric learning-based method for few-shot open-set keyword spotting. Using dummy prototypes, our approach effectively rejects unknown classes and outperforms recent FSOSR methods on both splitGSC and miniImageNet benchmarks.

We propose PK-MTL, a multi-task learning framework for personalized keyword spotting that jointly learns keyword spotting and speaker verification.

We present a cosine similarity-based method for unsupervised domain adaptation in semantic segmentation. By aligning source and target features via a class-wise feature dictionary and similarity maximization, our approach reduces domain gaps and improves performance on standard UDA benchmarks.

We present a real-time multi-modal framework for outdoor robots that detects, tracks, and classifies human actions using RGB, thermal, and LiDAR data.

We propose MDOD, a novel object detector that formulates multi-object detection as bounding box density estimation using a mixture model, simplifying training and improving performance.

We propose Feature Fusion Learning (FFL), a mutual knowledge distillation framework that combines features from diverse sub-networks to train a stronger fused classifier, improving both overall and individual network performance.

We propose an online knowledge distillation method that aligns feature maps via adversarial training, enabling networks to learn from each other's internal representations and improving performance, especially in mixed-size network setups.

Education

Ph.D. in Engineering, Seoul National University, 2019–2024
Graduate School of Convergence Science and Technology, Department of Intelligence and Information
Advisor: Nojun Kwak

B.S. in Creative Technology Management & Computer Science, Yonsei University, 2012–2019
Exchange Student Program at Technical University of Munich (2016–2017)

Work Experience

Staff Researcher, Samsung Electronics, 2024.Mar–Present
Working on Vision-Language-Action models for robotics, integrating visual, linguistic, and action-based inputs to enhance perception and decision-making in autonomous systems.

Research Intern, Qualcomm AI Research Korea, 2022.Jan–2022.Jul
Developed few-shot learning methods for keyword spotting to recognize new keywords from minimal examples, enhancing speech recognition performance.

Research Intern, Naver Webtoons Corp., 2020.Jun–2020.Dec
Researched unsupervised domain adaptation for semantic segmentation, improving model generalization across diverse webtoon styles through novel adaptation strategies.

Patents

Domestic Patents

• Device for Unsupervised Domain Adaptation in Semantic Segmentation Exploiting Inter-pixel Correlations and Driving Method Thereof


• Device for Regression Scale-aware Cross-domain Object Detection and Driving Method Thereof

International Patents

• Method and Apparatus with Online Task Planning


• Method and Apparatus with Micro-Action Determination


• Multi-Task Learning for For Personalized Keyword Spotting


• Dummy Prototypical Networks For Few-Shot Open-Set Keyword Spotting

Funding

• Samsung Electronics Ph.D. Student Sponsorship Program (2021-2024)

Honour & Awards

• The best paper award in the 3rd Workshop on Real-World Surveillance: Applications and Challenges at WACV2023, $1,000 (2023)


• BK21 Graduate School Innovation Project Colloquium Outstanding Graduate Student Award, ￦500,000 (2021)


• Qualcomm Innovation Fellowship Korea 2020 Final Session, Finalist (2020)


• Samsung Electronics - Seoul National University, Industry-Academic Cooperation Research Award, ￦1,500,000 Won (2020)


• Special Award, Yonsei Programming Contest 2018 (Spring Semester 2018)


• Honours, tuition support scholarship, Yonsei University (Spring Semester 2016)