Xudong Cai

I am a Ph.D. candidate at Renmin University of China (RUC), advised by Yongcai Wang and Deying Li, and a visiting student in the Physical Vision Group (PVG) at Nanyang Technological University (NTU), working with Chuanxia Zheng. My research focuses on 3D computer vision, with an emphasis on 3D scene reconstruction and generation; I welcome discussions and collaborations—feel free to reach out.

Email  /  Scholar  /  GitHub

Research

I mainly work on 3D computer vision, with an emphasis on 3D scene reconstruction and generation.

• Scene Generation
• 3D / 4D scene reconstruction
• Image matching
• Visual Place Relocalization

Mem4D decouples static geometry and dynamic motion by introducing persistent structure memory and transient dynamics memory, enabling high‑fidelity dynamic scene reconstruction with global consistency.

We propose a monocular vision‑language navigation framework that "dreams" panoramic context from partial observations to bridge the gap between limited field‑of‑view inputs and global spatial reasoning.

This work explores auxiliary reasoning tasks to improve data efficiency in vision‑language navigation, allowing agents to learn complex navigational strategies without massive datasets.

Dust to Tower introduces an accurate and efficient coarse‑to‑fine framework that jointly optimizes 3D Gaussian Splatting and image poses from sparse, uncalibrated images, enabling photo‑realistic scene reconstruction.

MambaVO integrates sequential matching refinement and a trending‑aware training penalty into a Mamba‑based architecture for efficient visual odometry, yielding robust pose estimation and mapping.

PRISM maximizes progressive dependency to explicitly model scale relationships, improving feature matching accuracy for images with extreme scale variations.

Voloc performs cross‑modal visual place recognition by querying a compressed LiDAR map with visual images, achieving robust localization in large‑scale environments with reduced storage.

Vsformer is a transformer‑based architecture that mines correlations within a flexible set of views to enhance multi‑view 3D shape classification and retrieval.

ColSLAM is a collaborative SLAM system designed for mobile phones that fuses point and line features and introduces a novel map caching mechanism to support robust multi‑agent mapping.

This work fuses semantic features and geometric structures from point clouds to improve 3D object detection accuracy in complex environments.

This Chinese‑language survey categorizes visual relocalization methods based on prior maps (2D, point‑cloud, etc.) and discusses current challenges and future directions.