Minha Lee

Computer Science & Engineering student with a passion for Computer Vision and Autonomous Robotics.

About Me

Hi, I’m Minha Lee, a student deeply interested in computer vision and robotics, especially 3D reconstruction and world model building. I hold a B.S. in Computer Science and Engineering from Korea University. I have conducted research in both computer vision and robotics across Korea University, the University of Washington, Korea Institute of Science and Technology, and Korea Advanced Institute of Science and Technology.

Research Interests

#Computer Vision | #3D Reconstruction | #Robotics | #World Model | #Deep Learning

Robust 3D Reconstruction: Developing methods that maintain reconstruction quality under occlusion, noise, and incomplete observations.
World Model Construction: Build object-centric and scene-level world models from accurate 3D reconstructions, enabling robots to understand, predict, simulate, and plan in their environments.
Automated Data Generation for Robotics: Use world-model-based simulations to generate high-quality training data, reducing reliance on costly real-world experiments.

Education

Korea University Mar 2022 - Feb 2026

B.S. in Computer Science and Engineering

Major GPA: 4.33/4.5 [Link]
Relevant Coursework: Data Structure, Algorithms, Deep Learning, Artificial Intelligence, Natural Language Processing, Introduction to Compute Vision Its Application, Operating Systems, Interactive Visualization, Machine Learning

University of Washington Mar 2025 - Jun 2025

Exchange Student in Computer Science and Engineering

Exchange GPA: 3.75/4.0 [Link]
Relevant Coursework: Autonomous Robotics, AI-Robotics, Research Seminar - Robotics Colloquium

Experience

Research Intern

Korea Advanced Institute of Science and Technology (KAIST) Seoul, Korea
Mar 2026 - Present

Advised by Prof. Jinwoo Shin

Conducting research on multi-view image-to-3D mesh reconstruction for robust 3D environment understanding.
Experimenting with reconstruction models to identify limitations and explore potential improvements.

Humanoid Research Center, Korea Institute of Science and Technology (KIST) Seoul, Korea
Sep 2025 - Dec 2025
Mar 2026 - Present

Advised by Dr. Jung-Min Park

Developed HCI (Human-Computer Interaction) and HRI (Human-Robot Interaction) technologies within Mixed Environments (VR, XR).
Conducted a survey and trend analysis in the field to summarize recent advances.
Developed a context-aware 3D Undo System for physics-based VR environments to address the irreversibility of cascading physical interactions.
- Engineered dual restoration mechanisms: an action-based approach and an object-based approach.
- Validated the system's efficacy through comprehensive user studies, analyzing task performance metrics and usability scores.
- Authored a research paper detailing the system architecture and user study findings.
Participated in the development of AI-driven robotic planning, with the goal of enabling robots to reason and generate task plans in a manner similar to human thinking.
Currently developing a robotics system integrating ROS, Unity, and database components, enabling a robot to manipulate valves in a simulated environment and preparing for deployment in real-world robotic platforms.

Personal Robotics Lab, University of Washington Seattle, Washington
Mar 2025 – Jan 2026

Advised by Prof. Siddhartha Srinivasa

Conducted research at the intersection of 3D vision and robotics, focusing on occlusion-robust object reconstruction for robotic manipulation.
Reviewed 3D Reconstruction literature and analyzed recent research trends.
Investigated failure cases of BundleSDF under hand-object occlusion and implemented reconstruction improvements across pre-processing and post-processing stages.
Built a custom RGB-D dataset using an Orbbec camera to evaluate reconstruction performance under real interaction scenarios.
Extended the project into a Real2Sim pipeline by integrating 3D mesh generation with physical parameter estimation (mass, center of mass, inertia), enabling the construction of simulation-ready digital twins.

Computer Vision Lab, Korea University Seoul, Korea
Jul 2024 - Dec 2024

Advised by Prof. Sangpil Kim

Conducted an in-depth analysis of literature in 3D vision and Novel View Synthesis (NVS).
Investigated the geometric limitations of feed-forward 3D Gaussian Splatting models (e.g., pixelSplat), identifying depth map fidelity as the primary bottleneck for NVS quality.
Leveraged single-view depth estimation models, such as Marigold and EVP, to generate high-fidelity depth maps, thereby enhancing the geometry accuracy of 3D reconstruction.

Hypermedia Lab, Korea University Seoul, Korea
Jan 2023 - Jul 2024

Advised by Prof. Seongbin Park

Reviewed and summarized research papers on P vs. NP and contributed to writing a related paper.

Projects

📍 : selected projects

📍 A Portable System for 3D Reconstruction and Physical Parameter Estimation of Grasped Objects Mar 2025 – Jan 2026

Undergraduate Research/Projects at the UW Personal Robotics Lab
Investigated structural reconstruction failures in BundleSDF, which struggled to restore missing geometry during 3D reconstruction under occlusion.
Designed a layered mitigation strategy spanning pre-processing, model-level exploration, and post-processing refinement.
- Pre-processing: Applied Inpaint-Anything to reduce occlusion-induced missing geometry prior to reconstruction.
- Model-level: Attempted to leverage category-level priors to guide the reconstruction of missing parts using AutoSDF.
- Post-processing: Integrated mesh filling or InstantMesh fusion to seamlessly complete the object surfaces.
Constructed a custom RGB-D dataset using an Orbbec camera to evaluate reconstruction performance under real interaction scenarios.
Extended the reconstruction pipeline into a Real2Sim framework by estimating physical parameters (mass, center of mass, inertia) and constructing simulation-ready digital twins.

Ensured generalizability by designing the system to rely only on data captured from the robot's gripper, making the solution robot-arm-agnostic.
Collaborated with another undergraduate researcher who managed the hardware setup for dataset collection.

Highlighted the limitations of purely geometric reconstruction and emphasized the integration of perception and dynamics for embodied world modeling.
Tools Used: Python, ROS, Ubuntu

📍 CUT3R-Based Surprise Measurement
in General 3D World for Robotic System Apr 2025 – Jun 2025

Term project for the AI-Robotics course of 2025-Spring quarter, resulting in a poster presentation.
Researched and implemented CUT3R for surprise measurement.
Collected and preprocessed real-world video datasets in static indoor and dynamic outdoor environments.
Quantified surprise in the CUT3R model by tracking changes in persistent spatial understanding via L-infinity norm and detecting events exceeding an EMA-based adaptive threshold.
Validated the framework on driving datasets, successfully detecting six key surprise events aligned with intuitive expectations.
Conducted an individual project extension, integrating Vision-Language Models (VLMs) to annotate datasets, and attempted to develop a framework for more robust and stabilized surprise measurement.
Tools Used: Python

Vision-Based Interior Designer Dec 2024 – Feb 2025

Team project conducted at a Deep Learning Society AIKU
Researched papers and code related to 3D reconstruction and similar projects.
Explored and implemented technologies such as DreamGaussian and InstantNeRF.
Generated depth maps using Marigold and created meshes from the depth maps.
Captured multi-view videos and applied GAN-based models with textual prompts, followed by 3D reconstruction for rendering and visualization.
Tools Used: Python

Predictive Modelling for Online Ad Click Behaviour Dec 2024 – Jan 2025

Term project for the Machine Learning course of 2024-Winter session
Preprocessed the dataset using one-hot encoding.
Trained various models including logistic regression, classification trees, and random forests.
Evaluated model performance using accuracy, precision, recall, and F1-score.
Tools Used: Python

Cafe Recommendation System Sep 2024 – Dec 2024

Team project conducted at a Deep Learning Society AIKU
Collected and structured a large-scale dataset of online cafe reviews and images via web crawling.
Implemented the core system using a cafe image classification model combined with prompt engineering and iterative refinement for model generation.
Tools Used: Python

Sketch-based Fashion Item Retrieval system (FashionSBIR) Jun 2024 – Aug 2024

Team project conducted at a Deep Learning Society AIKU
Preprocessed the dataset, including the addition of Canny edge maps.
Leveraged CLIP’s vision and text encoders to extract embeddings for images, sketches, and text, and trained a metric-learning model using contrastive learning.
Tools Used: Python

Technical Science Expert Question and Answer Model Project Mar 2024 – May 2024

Term project for the Natural Language Processing course of 2024-Spring semester
Developed a question-answering model using BERT, Longformer, and BART architectures.
Preprocessed domain-specific documents into SQuAD format, creating multiple datasets with or without HTML tags.
Conducted comparative experiments across models and preprocessing variations, evaluating performance with custom F1 and Exact Match (EM) metrics to validate hypotheses and analyze model sensitivity.
Tools Used: Python, PyTorch, Hugging Face

Korean Sign Language Translator Mar 2024 – Jun 2024

Term project for the Introduction to Computer Vision Its Application course of 2024-Spring semester
Developed a real-time Korean Sign Language translator using ViViT-based model.
Preprocessed a sign language video dataset by extracting joint coordinates with MediaPipe and overlaying them on the videos.
Developed a dictionary-like module for storing word feature representations.
Achieved best training performance by fine-tuning on the pose-landmark overlay dataset, with a validation loss of 0.8679 and Top-1 accuracy of 87.0%.
At inference, combining Manhattan Distance with an RBF kernel and average pooling on the original dataset achieved the highest accuracy of 57.0%.
Tools Used: Python, PyTorch, Hugging Face

Publications

[2] Minha Lee, Jung-Min Park, Sang-Yup Han. Toward Context-Aware and Causally Coherent Undo in Physics-based Virtual Reality, Accepted for presentation and publication (to appear) in the Proceedings of HCI International 2026 (HCII 2026) – Affiliated Conference: Virtual, Augmented and Mixed Reality (VAMR 2026), Springer, 2026 [Link]

Presentations

[1] Minha Lee, Soyeon Jeong, and Seongbin Park. On hypotheses under which P = NP, Informal Presentation at Computability in Europe 2023 (CiE 2023), July 24-28, 2023 [Link]

Skills & Technologies

Programming Languages: Python, C, SQL, C++, HTML, ROS
Frameworks & Libraries: PyTorch, TensorFlow, OpenCV, NumPy, SciPy
Tools & Platforms: GitHub, Linux, VS Code, Jupyter Notebook, Hugging Face, Unity
English Skills: TOEFL ibt 105 (Mar 2025) [Link]

Additional Information

Extracurricular Activities

KUCC: Computer Club, Korea University Sep 2022 – Mar 2023
- Interdisciplinary Community: As a central university club, engaged with members from a wide range of academic backgrounds.
- Django Study: Participated in weekly Django study sessions, following textbooks and coding along to strengthen practical skills.
KWEB: Web Development Club, Korea University Mar 2023 – Feb 2024
- Web Development Study: Conducted studies on web development topics, including HTML, databases, and SQL. Held once a week.
AIKU: Deep Learning Society, Korea University Dec 2023 - Feb 2025
- Deep Learning Study: Reviewed fundamental concepts of deep learning and explored various models, with a focus on studying the latest advancements in the field. Held twice a week.
- Paper Study: Selected and presented academic papers weekly, leading discussions to share insights, ask questions, and explore different perspectives. Held once a week.
- Project: Collaborated in teams to work on various deep learning-related projects, tackling different topics and challenges. Completed one project each semester and during vacation periods.