CVIA has led or contributed to multiple major national research tasks, including the Chinese Lunar Exploration Program, the manned space program, deep-space exploration missions, and the 10,000-meter Mariana Trench deep-sea expedition. The group has also deployed key technologies and systems in remote-sensing satellites, aerial refueling, and ground unmanned systems.
In addition, CVIA actively participates in academic and industrial algorithm challenges at both the international and domestic levels, with strong competition results. Most algorithms that are free of proprietary restrictions have been open-sourced for broader academic and industrial discussion.
Selected publications from the lab.
Published at CVPR 2025. The paper proposes a plug-and-play pose refinement algorithm that injects shape-constrained scene flow into the optimization loop, improving pose accuracy and scene adaptability on top of an existing initial pose. The method can be cascaded with existing pose-estimation pipelines.
Published at ICCV 2023. The paper presents a self-supervised 6D pose-estimation framework that does not require depth or extra annotations, and improves generalization and refinement in real scenes through pseudo-flow supervision derived from multi-view geometric consistency.
Published at CVPR 2023. The work implicitly embeds 3D object shape into iterative matching and pose solving, delivering more efficient and accurate 6D pose refinement under occlusion and large viewpoint changes.
Published in IEEE Transactions on Multimedia. This work proposes a structure-aware graph convolution network for point-cloud parsing, with adaptive dilated neighborhood search, learnable graph filtering, and structure-aware feature transformation.
Published in ISPRS Journal of Photogrammetry and Remote Sensing. The work introduces an efficient hierarchical framework for satellite attitude determination and improves attitude estimation accuracy.
Published at ACM Multimedia 2021. This paper achieved leading benchmark performance in point-cloud completion and is one of the group’s representative contributions in 3D geometric learning.
Representative deployment of the lab’s research in major engineering projects.
For high-fidelity transmission of lunar-surface operation videos captured by the lander robotic arm, the team developed a multi-channel, multi-mode real-time video encoding unit for the ascender sampling and separation-monitoring cameras. The system supports real-time processing of five camera streams under different resolutions, operating modes, compression formats, and analysis needs.
To support real-time video monitoring and guidance during rendezvous and docking between the space station and cargo spacecraft, the team developed a multi-channel HD video communication and image-processing system for mission monitoring and ground analysis.
For live broadcasting during deep-sea trials in the Mariana Trench, the team developed a multi-channel cinema-grade 4K ultra-HD encoding system together with ship-side decoding equipment, achieving stable operation in harsh deep-sea conditions.
For attitude perception and early warning of complex space targets, the group built an intelligent real-time perception unit based on target detection, pose estimation, and pose tracking for SDA applications.
To meet the need for precise localization in aerial refueling, the team developed a prototype system integrating visual perception, target measurement, and pose estimation, enabling high-precision 3D geometric perception in aviation scenarios.
Graduate students from the group have participated in multiple international and domestic algorithm challenges, benchmark competitions, and industry contests, and achieved excellent results.
A joint team from Xidian University, EPFL, and Magic Leap won first place in the single-model track of the BOP Challenge at ICCV 2023 in Paris and was invited to present at the 8th International Workshop on Recovering 6D Object Pose (R6D).
A joint team from Xidian University, EPFL, and Magic Leap won first place in the single-model track of the BOP Challenge at ECCV 2022 and was invited to report at the 6th International Workshop on Recovering 6D Object Pose.
The group won third place in the NTIRE 2022 Challenge on Spectral Recovery at the CVPR 2022 Workshop, and later moved up to second place in the post-challenge generalization test.
ASFM-Net ranked first on Stanford’s Completion3D leaderboard in October 2021 among 26 competing methods.
The team won the Track 1 Clean Winner Award in the NTIRE 2020 Challenge on Spectral Reconstruction from an RGB Image at the CVPR 2020 Workshop.
The group won first place in Subject II of the 2019 Tianzhi Cup AI Challenge in the surveying, mapping, meteorology, and hydrology domain.
The group won second place in Subject I, “3D framework construction for typical objects,” based on multi-view satellite imagery.
