Minsik Jeon

Hello! I am a MS in Robotics student at Carnegie Mellon University (CMU), advised by Prof. Shubham Tulsiani. I am interested in 3D and 4D reconstruction from perceptual input, and its applications in robotics.

Prior to joining CMU, I served as a researcher at the Agency for Defense Development (ADD) in South Korea while concurrently completing my mandatory military service. I completed my bachelor's degree (Summa Cum Laude) in School of Computing at KAIST (Korea Advanced Institute of Science and Technology).

Email CV Scholar LinkedIn X

Latest News

Mar 2026

One paper accepted to CVPR 2026! Flow3r is a factored flow-prediction framework for scalable 3D/4D reconstruction, trained on ~800K unlabeled videos.

Jan 2026

One paper accepted to RA-L! E2-BKI is an uncertainty-aware Gaussian semantic mapping framework that combines Evidential Deep Learning with Bayesian Kernel Inference for outdoor robots.

Nov 2025

One paper accepted to WACV 2026! OWOD-Rep is an Open-World object detection method that leverages Vision Foundation Models for novel-object detection and semantically rich feature extraction.

Aug 2025

Started M.S. in Robotics at Carnegie Mellon University, advised by Prof. Shubham Tulsiani.

May 2024

Presented DA-RAW at ICRA 2024 in Yokohama, Japan.

Publications (* denotes equal contribution)

Point4D: Long-range 4D Motion Reconstruction

Under submission

Preprint, 2026

PDF

A feed-forward model for long-range 4D motion reconstruction that infers dense per-point 3D trajectories across hundreds of frames. By decoding motion for 3D query points instead of 2D pixels, predicted endpoints are directly re-queried across overlapping video chunks, enabling robust long-range tracking under occlusion.

RayRoPE: Projective Ray Positional Encoding for Multi-view Attention

Yu Wu, Minsik Jeon, Jen-Hao Rick Chang, Oncel Tuzel, Shubham Tulsiani

Preprint, 2026

Project arXiv Code

A 3D-aware positional encoding for multi-view transformers that uses predicted depths along rays and projective coordinates to achieve SE(3)-invariant attention adapted to scene geometry.

Novel View Synthesis as Video Completion

Qi Wu, Khiem Vuong, Minsik Jeon, Srinivasa Narasimhan, Deva Ramanan

Preprint, 2026

Project arXiv

Reformulates sparse novel view synthesis as low frame-rate video completion, adapting video diffusion models to unordered multi-view inputs via per-frame latent encodings and removal of temporal positional embeddings.

Flow3r: Factored Flow Prediction for Scalable Visual Geometry Learning

Zhongxiao Cong, Qitao Zhao, Minsik Jeon, Shubham Tulsiani

IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026

Project arXiv Code Demo

A scalable framework for 3D/4D reconstruction that uses dense 2D correspondences from unlabeled videos as supervision. Factored flow prediction decomposes flow into geometry and pose components, enabling training on ~800K unlabeled videos.

Open-World Object Detection with Instance Representation Learning

Sunoh Lee*, Minsik Jeon*, Jihong Min, Junwon Seo

IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2026

Project Video arXiv

A method for Open-World object detection that improves novel object detection and semantically rich feature extraction by leveraging Vision Foundation Models.

E2-BKI: Evidential Ellipsoidal Bayesian Kernel Inference for Uncertainty-aware Gaussian Semantic Mapping

Junyoung Kim, Minsik Jeon, Jihong Min, Kiho Kwak, Junwon Seo

IEEE Robotics and Automation Letters (RA-L), 2026

arXiv PDF

An uncertainty-aware semantic mapping framework that combines Evidential Deep Learning with Bayesian Kernel Inference, capturing multiple sources of uncertainty for outdoor robot mapping while preserving real-time efficiency.

DA-RAW: Domain Adaptive Object Detection for Real-World Adverse Weather Conditions

Minsik Jeon*, Junwon Seo*, Jihong Min

IEEE International Conference on Robotics and Automation (ICRA), 2024

Project Video arXiv

An UDA framework for object detection that can effectively adapt to real-world adverse weather conditions by addressing Style & Weather gaps separately.

Projects

Adaptive Path Planning Based on Situational Awareness and Dynamic Model Learning

Agency for Defense Development, 2023 – Present

Developing perception and control algorithms for off-road navigation: robust moving object detection & tracking in off-road environments, and semantic terrain classification maps via LiDAR–Camera integration. Examined path planning and control algorithms for navigation in complex environments.

Unmanned Reconnaissance Vehicles Development

Agency for Defense Development, 2023

Developed a LiDAR–IR fusion method for robust object detection under visibility-constrained scenarios.

Deformable Object Recognition Technology

Agency for Defense Development, 2022 – 2023

Developed an open-set 2D & 3D object detection method for LiDAR pointcloud.

Indy Autonomous Challenge (IAC)

USRG @ KAIST (Advised by Prof. David Hyunchul Shim), 2021

Participated in the Indy Autonomous Challenge (IAC), the first autonomous car racing competition to be held in Indianapolis, as an intern for Team KAIST. As part of the perception team, I tested algorithms using camera and LiDAR in a simulation environment.

Microscopy Image Artifact Removal & Super-resolution

NICA Lab @ KAIST (Advised by Prof. Young-Gyu Yoon), 2021

Developed a removal network that addresses the line artifact of SCAPE microscopy images, using a CycleGAN-like model and transferring the image into a CFM style.

Performance and Operation Analysis of On-Board RAID

SK Hynix, 2020

In the Performance Analysis team, configured On-board RAID using SSDs and conducted performance evaluations for each RAID option.