Professor and PhD advisor. His research and applications center on 3D spatial intelligence, 3D localization and measurement, 6D pose estimation, and engineering vision systems.
He received the B.S., M.S., and Ph.D. degrees from Xidian University in 2003, 2006, and 2009, respectively, and was a visiting researcher at the University of Southern California from 2012 to 2013.
He is a member of the ISN State Key Laboratory and the Institute of Image Transmission and Processing, a member of IEEE, a council member of the Shaanxi Society of Image and Graphics, and a committee member of both the CCF-CV and CSIG-3DV technical committees. He also serves as a reviewer for IEEE TPAMI, TNNLS, TMM, TCSVT, TGRS, TIP, Pattern Recognition, and Remote Sensing.
He leads the Shaanxi University Youth Innovation Team and serves as chief scientist of the Shaanxi “Scientist + Engineer” team.
His honors include individual commendation for Chang’e-5, a first-class Ministry of Education Science and Technology Progress Award, a second-class Shaanxi Science and Technology Award, a first-class Surveying and Mapping Science and Technology Award, a second-class AVIC Science and Technology Progress Award, the Shaanxi Rising Science Star award, and Xidian University Huashan Scholar recognition.
His main research interests include 3D spatial intelligence, 3D localization and 6D pose estimation in computer vision, as well as intelligent processing of remote sensing imagery.
His research follows the principle that problems should be distilled from engineering practice and results should be validated in real applications, with continuous closed-loop advancement from algorithm design to hardware implementation and system deployment.
His major research directions are 3D spatial intelligence, 3D localization, and 6D pose estimation for computer vision. In recent years he has built a continuous line of work on geometric perception of complex space targets, target pose estimation, structure-constrained modeling, point-cloud understanding, and deployment of engineering vision systems, while also covering key algorithms and architectures for image/video compression and transmission.
Over the past five years, he has led multiple government-funded research projects and more than ten industry-commissioned R&D tasks from institutes affiliated with CASC, AVIC, CAS, NORINCO, and CETC. During the same period, he has published more than 60 SCI papers, including more than 40 as first or corresponding author, with over 50 papers in CAS Q2 or above, together with multiple national invention patents.
For high-fidelity downlink of lunar-surface operation videos captured by the lander robotic arm, the group developed a real-time multi-channel, multi-mode encoding unit for the ascender sampling and separation-monitoring cameras. The system supports five camera streams with different resolutions, operating modes, compression formats, and analysis requirements.
To support real-time video monitoring and guidance during rendezvous and docking between the space station and cargo spacecraft, the group developed a multi-channel HD video communication and image-processing system for stable mission support and ground analysis.
For live broadcasting during Mariana Trench deep-sea trials, the group developed a multi-channel cinema-grade 4K encoding system together with ship-side decoding equipment, achieving stable performance in a harsh deep-sea environment.
For attitude perception and early warning of complex space targets, the group built a real-time intelligent perception unit based on target detection, pose estimation, and pose tracking, enabling more autonomous space-situation warning capability.
To meet the need for precise localization in aerial refueling, the group developed an integrated system for visual perception, target measurement, and motion guidance, supporting high-precision relative pose estimation under complex flight conditions.
The group studied key techniques for 2D–3D registration and real-time DSM projection to fuse UAV surveillance imagery with geospatial data. Pixels in the UAV field of view can be projected directly onto a 3D DSM, improving real-time monitoring, localization, and situational understanding.
A joint team from Xidian University, EPFL, and Magic Leap won first place in the single-model track of the BOP Challenge and was invited to present at the R6D workshop. The winning method followed a detection–estimation–refinement pipeline and led clearly in both RGB and RGB-D settings.
A joint team from Xidian University, EPFL, and Magic Leap won first place in the single-model track of the BOP Challenge and was invited to present at the corresponding international workshop.
ASFM-Net ranked first on Stanford’s Completion3D leaderboard and became one of the group’s representative achievements in point-cloud completion.
The team reconstructed 3D structures of target regions from multi-view satellite imagery, demonstrating strong algorithmic capability for demanding engineering tasks.
This paper introduces deformation modeling for 6D pose estimation under non-rigid deformation and heavy occlusion, improving robustness and estimation accuracy while preserving geometric consistency.
We propose a plug-and-play pose refinement algorithm that injects shape-constrained scene flow into the optimization loop, improving both estimation accuracy and scene adaptability from an existing initial pose.
This work presents a self-supervised 6D pose-estimation framework without depth or additional annotations, and improves real-scene generalization and refinement via pseudo-flow supervision derived from multi-view geometric consistency.
The paper proposes a shape-constrained recurrent-flow framework that implicitly embeds 3D object shape into iterative matching and pose solving for more efficient and accurate 6D pose refinement.
For 6D pose estimation under complex occlusion, the paper proposes a rigidity-aware detection strategy that uses visibility-guided sampling to obtain more stable detection initialization.
This work proposes SA-GCN, a structure-aware graph convolution network with adaptive dilated KNN, learnable graph filters, and structure-aware feature transformation for point-cloud classification and segmentation.
For remote-sensing satellite attitude determination, the paper proposes an efficient hierarchical framework combining adaptive EKF, ELM-based error compensation, and weighted smoothing to improve high-precision attitude estimation.
For onboard processors with limited computational capability, the paper proposes an ultra-lightweight satellite centroid localization method based on line and contour features, enabling accurate and real-time localization at low cost.
This work proposes a boundary-aware self-supervised pretraining framework for point-cloud semantic segmentation, improving boundary-region and fine-structure segmentation quality.
For real-time on-orbit detection of space targets against deep-space backgrounds, the paper proposes a target-detection algorithm based on morphological processing and inter-frame matching, validated on FPGA and DSP platforms.
ASFM-Net improves point-cloud completion by learning prior knowledge in feature space through an asymmetrical Siamese feature-matching autoencoder.
A robust interpolation framework for large-displacement optical flow improves dense flow recovery from sparse but reliable correspondences.
This work develops an efficient coarse-to-fine PatchMatch strategy for large-displacement optical flow, balancing accuracy and efficiency.
The entries below are uniformly formatted in an IEEE Transactions-style reference format. Some papers are still in early access or online-ahead-of-print status, so formal volume, issue, and page information may not yet be available and is therefore kept as early-publication metadata.
Goal: The course introduces how higher mathematics, matrices, and related knowledge are applied in real engineering problems, and trains students to translate engineering problems into mathematical formulations and solve them approximately with optimization methods. It also serves as a foundation for later study in machine learning, data mining, and deep learning.
Goal: The course covers basic concepts and architectures of computer communication networks, protocol mechanisms and key protocols, basic analytical methods for communication problems, and recent advances in networking technology, with emphasis on system understanding and engineering capability.
CVIA recommends LaTeX for master’s theses to reduce time wasted on formatting and separate content from layout. The Graduate School of Xidian University provides an official LaTeX template, and alumni from the institute have also fixed several issues in that template.
An undergraduate student from the 2020 class shared an undergraduate thesis template on GitHub. Students working on their bachelor thesis are encouraged to start from that template.
One additional academic master’s opening for 2026: Interested students are welcome to contact us by email. Replies are usually fast during working hours, and an online meeting can be arranged.
PhD openings are still available for 2026: Students interested in computer vision, image analysis, 3D spatial measurement, and 6D pose estimation are welcome to get in touch.
