Zhenjun Zhao (赵祯俊)
I am a postdoctoral researcher at University of Zaragoza, working with Javier Civera.
I received my PhD from the Chinese University of Hong Kong (CUHK), under the supervision of Ben M. Chen.
My research focuses on certifiable global optimization and robust estimation for 3D geometric perception, aiming to develop solvers that are globally optimal, outlier-robust, and computationally efficient. I work on problems across camera geometry estimation, Structure-from-Motion, SLAM, and visual localization, with a particular interest in bridging the gap between optimization theory and real-world robotic perception. My long-term goal is to build trustworthy geometric perception systems that provide strong correctness guarantees while remaining practical for deployment in complex, dynamic environments.
I am on the job market in 2026–2027. Please feel free to reach out!
News
- 10/2026 Grant: I am awarded a Marie Skłodowska-Curie Actions Postdoctoral Fellowship (MSCA-PF 2025) by the European Commission for the project TRUST-VSLAM at the University of Zaragoza!
- 10/2026 Invited to give a talk on Trustworthy Geometric Perception at Fauna, Amazon.
- 09/2026 Invited to give a talk on Trustworthy Geometric Perception at Eastern Institute of Technology, Ningbo.
- 07/2026 Invited to give a talk on Trustworthy Geometric Perception at Best of ICRA.
- 06/2026 Our paper MyGO-Splat is accepted to IROS 2026! Our paper Skyfall-GS is accepted to ECCV 2026!
- 06/2026 Our paper SGS receives the Best Paper Award at SPAR-3D Workshop @ CVPR 2026!
- 04/2026 Our paper ULF-Loc is selected as Highlight by CVPR 2026!
- 03/2026 Our paper Global Solvers for 3D Vision is accepted to ICRA 2026 Workshop!
- 02/2026 Our paper ULF-Loc is accepted to CVPR 2026! Our paper AdaGaR is accepted to CVPR Findings 2026!
- 01/2026 Our paper BlurPoint is accepted to ICRA 2026!
- 01/2026 Our paper NPC is accepted to ICLR 2026!
Research
Skyfall-GS: Synthesizing Immersive 3D Urban Scenes from Satellite Imagery
European Conference on Computer Vision (ECCV), 2026
A framework that synthesizes photorealistic, city-block scale 3D urban scenes from satellite imagery using diffusion models, eliminating the need for costly 3D annotations while enabling real-time exploration.
Global Solvers for 3D Vision: Foundations, Frontiers, and a Call to the Robotics Community
ICRA 2026 Workshop on Frontiers of Optimization for Robotics, 2026
A position paper presenting the landscape of global solvers for 3D vision, identifying open challenges in scalability, deep learning integration, and standardized evaluation, with a call to the robotics optimization community.
ULF-Loc: Unbiased Landmark Feature for Robust Visual Localization with 3D Gaussian Splatting
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026
· Highlight (top 3.6%)
· (* equal contribution)
A bias-free visual localization framework that replaces α-blending feature optimization with geometry-weighted multi-view fusion for accurate 2D-3D matching with 3D Gaussian Splatting.
AdaGaR: Adaptive Gabor Representation for Dynamic Scene Reconstruction
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Findings, 2026
A unified framework for reconstructing dynamic 3D scenes from monocular video by combining Adaptive Gabor Representation with Cubic Hermite Splines and temporal curvature regularization.
BlurPoint: Efficient Motion Blur Aware Student-Teacher Local Feature Learning
IEEE International Conference on Robotics and Automation (ICRA), 2026
· († corresponding author)
An end-to-end student-teacher framework that jointly learns blur-aware local feature detection and description from blurred images in a self-supervised manner.
Neural Predictor-Corrector: Solving Homotopy Problems with Reinforcement Learning
International Conference on Learning Representations (ICLR), 2026
· (* equal contribution; † project lead)
The first unified framework for homotopy methods across robust optimization, global optimization, polynomial root-finding, and sampling, using reinforcement learning to automatically learn efficient solver policies.
Surgical Gaussian Surfels: Highly Accurate Real-time Surgical Scene Rendering using Gaussian Surfels
IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2026
· Best Paper Award @ SPAR-3D Workshop, CVPR 2026
A point-primitive-based rendering technique that combines surfels with 3D Gaussian Splatting to achieve high-fidelity reconstruction of dynamic surgical scenes.
SLAM-X: Generalizable Dynamic Removal for NeRF and Gaussian Splatting SLAM
ACM International Conference on Multimedia (ACM MM), 2025
· (* equal contribution)
A plug-and-play module for dynamic scene handling across diverse SLAM architectures.
TurboReg: TurboClique for Robust and Efficient Point Cloud Registration
IEEE/CVF International Conference on Computer Vision (ICCV), 2025
A highly efficient and robust point cloud registration method, supporting both CPU and GPU platforms.
STG-Avatar: Animatable Human Avatars via Spacetime Gaussian
IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2025
· (* equal contribution)
3DGS-based framework for high-fidelity animatable human avatar reconstruction.
Convex Relaxation for Robust Vanishing Point Estimation in Manhattan World
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025
· Oral & Best Paper Award Candidate (top 0.1%)
· (* equal contribution)
A globally optimal and outlier-robust method for vanishing point estimation in a Manhattan world, using convex relaxation techniques.
HeMoRa: Unsupervised Heuristic Consensus Sampling for Robust Point Cloud Registration
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025
A learnable sampling probability distribution for matches in robust estimation, supervision-free and reinforcement-inspired.
GlobalPointer: Large-Scale Plane Adjustment with Bi-Convex Relaxation
European Conference on Computer Vision (ECCV), 2024
· (* equal contribution)
A globally optimal and efficient large-scale plane adjustment, using alternating minimization and convex relaxation techniques.
DeDoDe v2: Analyzing and Improving the DeDoDe Keypoint Detector
IEEE/CVF CVPR Workshop on Image Matching: Local Features & Beyond, 2024
An improved keypoint detector built on top of DeDoDe.
BALF: Simple and Efficient Blur Aware Local Feature Detector
IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2024
A simple yet efficient and effective motion blur aware local feature detector.
Benchmark for Evaluating Initialization of Visual-Inertial Odometry
Chinese Control Conference (CCC), 2023
A novel benchmark for the evaluation of visual-inertial odometry (VIO) initialization.
Hong Kong World: Leveraging Structural Regularity for Line-based SLAM
IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2023
A novel structural model to describe structured scenes with vertical, horizontal, and sloping dominant directions.
SyreaNet: A Physically Guided Underwater Image Enhancement Framework Integrating Synthetic and Real Images
IEEE International Conference on Robotics and Automation (ICRA), 2023
A novel UIE framework combining synthetic and real data guided by the revised underwater image formation model and domain adaptation strategies.
Teaching
- Co-supervisor, Undergraduate Final Year Project, 2019–2020
- Co-supervisor, MSc Project, 2019–2020
- Teaching Assistant, Multivariable Calculus for Engineers, Spring 2021
- Teaching Assistant, Complex Variables for Engineers, Fall 2020
- Teaching Assistant, Probability and Statistics for Engineers, Spring 2020
- Teaching Assistant (Lead), Introduction to Control Systems, Fall 2019
Talks
-
Trustworthy Geometric Perception: Certifiable Global Optimization and Robust Estimation. Slides can be found here.
- Fauna, Amazon, October 5, 2026
- Eastern Institute of Technology, Ningbo, September 3, 2026
- Best of ICRA Virtual Meetup, July 21, 2026
Academic Services
- Conference Reviewer: CVPR, ICCV, ECCV, ICML, NeurIPS, ICLR, ACM MM, AAAI, ICRA, IROS, BMVC, 3DV, WACV
- Journal Reviewer: TPAMI, IJCV, IJRR, TRO