Gyeongsu (Bob) Cho

I am a Ph.D. student in the Artificial Intelligence Graduate School at UNIST, where I am a member of the 3D Vision & Robotics Lab under Prof. Kyungdon Joo's supervision. I received a B.Eng. in Mechanical Engineering from Chung-Ang University.

I am actively seeking internship opportunities at academic or industrial research teams.
I’m always open to new collaborations — feel free to reach out to me at threedv@unist.ac.kr.

Email / CV / LinkedIn / GitHub

Research

I am passionate about creating models that do not yet exist — models capable of reconstructing, animating, and generating dynamic 3D and 4D content, especially for animals and humans. My work combines synthetic data pipelines, parametric animal/human models, and generative methods to lift 2D images and videos into animatable 3D/4D representations. I enjoy building tools and models that unlock new creative and scientific possibilities in computer vision and graphics.

	4D Animal Mesh Reconstruction from Monocular Videos In submission (CVPR) [paper] / [project page] We present a synthetic animal video pipeline and a video transformer model for reconstructing temporally coherent 4D animal motion and global trajectories from monocular in-the-wild videos.
	Video Anomaly Detection in Display Inspection In submission (TCSVT) [paper] / [project page] We present a novel framework for video anomaly detection, where multiple devices are concurrently monitored by an external camera during operation. This work was done in collaboration with Samsung Electronics.
	Pose-Diverse Multi-View Virtual Try-on from a Single Frontal Image Seonghee Han, Minchang Chung, Gyeongsu Cho, Kyungdon Joo, Taehwan Kim WACV 2026 [paper] / [project page] This work generates pose-diverse, multi-view images for virtual try-on starting from a single frontal input. By leveraging multi-view consistency and pose conditioning, the method produces realistic, view-consistent outfits under large pose variations.
	Avatar++: Fast and Pose-Controllable 3D Human Avatar Generation from a Single Image Seonghee Han, Minchang Chung, Gyeongsu Cho, Kyungdon Joo, Taehwan Kim ICCV 2025, Wild3D Workshop [paper] / [project page] Avatar++ generates an animation-ready 3D human avatar from a single image in seconds. By combining identity-preserving features with pose-guided multi-view synthesis, the framework enables fast and controllable avatar creation for downstream animation and XR applications.
	DogRecon: Canine Prior-Guided Animatable 3D Gaussian Dog Reconstruction From a Single Image Gyeongsu Cho, Changwoo Kang, Donghyeon Soon, Kyungdon Joo International Journal of Computer Vision (IJCV), 2025 [paper] / [project page] DogRecon is a framework that reconstructs animatable 3D Gaussian dog models from a single RGB image. It leverages a canine prior and a reliable sampling strategy to obtain high-quality 3D shapes and realistic animations of dogs from in-the-wild photos.
	Canine Prior-Guided Animatable 3D Gaussian Dog Reconstruction From a Single Image Gyeongsu Cho, Changwoo Kang, Donghyeon Soon, Kyungdon Joo CV4Animals Workshop @ CVPRW 2025 (Oral) [paper] / [oral video] This workshop version presents the initial DogRecon framework for animatable 3D Gaussian dog reconstruction from a single image, focusing on canine priors and practical reconstruction quality for in-the-wild photos.
	SlaBins: Fisheye Depth Estimation using Slanted Bins on Road Environments Jongsung Lee, Gyeongsu Cho, Jeongin Park, Kyongjun Kim, Seongoh Lee, Jung Hee Kim, Seong-Gyun Jeong, Kyungdon Joo ICCV 2023 [paper] / [project page] SlaBins introduces a slanted multi-cylindrical representation and adaptive depth bins to achieve accurate and dense depth estimation from automotive fisheye cameras in road environments. This work was done in collaboration with 42dot.

Education

Integrated M.S.–Ph.D. in Artificial Intelligence, Artificial Intelligence Graduate School, UNIST, South Korea (2022.03 – Present)
B.Eng. in Mechanical Engineering, Chung-Ang University, South Korea (2018.03 – 2021.08)

Research & Industry Experience

Graduate Researcher, UNIST – 3D/4D Vision & Graphics (2022 – Present)
Lead author of DogRecon (IJCV 2025) and a 4D animal reconstruction framework for single-image 3D Gaussian dog reconstruction and 4D animal motion from monocular videos. Develop synthetic video pipelines, SMAL-based animal models, and PyTorch3D-based renderers for large-scale training and evaluation. Explore generative pipelines that lift text/video outputs into animatable 3D/4D representations.
Research Collaboration with Samsung Electronics – Video Anomaly Detection (2024 – 2025)
Co-developed a video anomaly detection framework for industrial display inspection using real manufacturing data. Contributed to dataset design, model architecture, and experiments for a TCSVT submission on reference-guided video anomaly detection.
Startup Project – Real-Time 3D Human Pose Estimation (KIC & GWU Program, 2024)
Built a 3-camera real-time 3D human pose estimation system with smoothing algorithms for stable motion analysis in gyms, designed as an AI-based personal training assistant. Selected for a Korean government startup support program and a 3-week entrepreneurship program at George Washington University. Conducted 50+ interviews with fitness and healthcare professionals across major U.S. cities and refined a US-market–focused business model, ranking 1st in track.
Other industry-funded projects
Monocular depth estimation (funded by 42dot, 2022), a recommendation system for virtual tactile stiffness (funded by UNIST, 2023), and human pose estimation (funded by NIA, 2021).

Template from this website.