Kezheng Xiong

Kezheng Xiong

he/him

3D computer vision researcher, a Ph.D. student in Xiamen University

Kezheng Xiong, Ph.D. Student at Xiamen University

About Me

I am a Ph.D. student in the Department of Artificial Intelligence at Xiamen University, where I also completed my undergraduate degree in Computer Science and Technology in 2024. I am affiliated with the ASC Laboratory at Xiamen University, where I conduct cutting-edge research in artificial intelligence, machine learning, and computer vision.

My research focuses on 3D computer vision, with particular emphasis on point cloud registration—a fundamental task in 3D computer vision. Since 2023, I have contributed to the field through research projects that advance the state-of-the-art in this domain. My work has been published in top-tier AI conferences, including AAAI, NeurIPS, CVPR, and ECCV.

Education

  • Ph.D. in Artificial Intelligence, Xiamen University, 2024 - Present
  • B.Sc. in Computer Science and Technology, Xiamen University, 2020 - 2024

For more details, please refer to my CV.

News

  • [06/2026]: One paper on unsupervised point cloud registration has been accepted to ECCV 2026.

  • [02/2026]: One co-authored paper has been accepted to CVPR 2026.

  • [09/2024]: One paper on unsupervised point cloud registration has been accepted to NeurIPS 2024.
  • [12/2023]: One paper on cross-source point cloud registration has been accepted to AAAI 2024.

Research Interests

My research focuses on the following areas:

  • Point Cloud Registration: Developing robust algorithms for aligning 3D point clouds, including cross-source and unsupervised settings.
  • Joint LiDAR Localization and 3D Object Detection: Unifying localization and perception tasks within a single contrastive learning framework.

Point Cloud Registration

Point cloud registration is a fundamental task in 3D computer vision that aims to align two or more point clouds into a common coordinate system. This technology has widespread applications in 3D reconstruction, object recognition, and robot navigation. The objective is to estimate the transformation \(\mathbf{T}= {\mathbf{R}, \mathbf{t}}\) that optimally aligns the source point cloud \(\mathbf{P}_s\) to the target point cloud \(\mathbf{P}_t\):

\[\mathbf{T} = \arg\min_{\mathbf{T}} \sum_{i=1}^{N} \left\| \mathbf{R}\mathbf{p}_{s_i} + \mathbf{t} - \mathbf{p}_{t_i} \right\|^2\]

where \(\mathbf{R}\) and \(\mathbf{t}\) represent the rotation matrix and translation vector, respectively.

My research addresses the following challenging scenarios:

  • Cross-source Point Cloud Registration (AAAI 2024): Aligning point clouds captured from heterogeneous sources, such as LiDAR and RGB-D cameras.
  • Unsupervised Point Cloud Registration
    1. (NeurIPS 2024): Developing registration methods without relying on ground-truth pose supervision during training.
    2. (ECCV 2026): Identifying a key failure mode of existing unsupervised methods and proposing a novel teacher-student framework with train-only semantic guidance to address it.

Other Research Participated

Beyond point cloud registration, I have also contributed to related areas in 3D computer vision:

  • Joint LiDAR Localization and 3D Object Detection (CVPR 2026): Contributed to TACO, the first task-aware contrastive learning framework that unifies LiDAR-based localization and 3D object detection within a single network. TACO leverages contrastive learning to explicitly decouple static geographic features for localization and object-centric features for detection, enabling bidirectional mutual supervision. This work also introduced the OxfoLD dataset, the first multi-traversal LiDAR dataset with both localization ground truth and rich 3D object annotations across various times and weather conditions.