Fei Xue

I am a PhD student in the Computer Vision Group at the University of Cambridge, supervised by Prof. Roberto Cipolla and Dr. Ignas Budvytis.

My research interests lie in large-scale 3D reconstruction and localization for autonomous driving, robotics, and AR/VR. I am especially interested in efficient and accurate pose estimation by introducing multimodality signas, semantics, and geometric constraints.

Prior to my PhD, I obtained the Bachelor and Master degrees from Peking University under the supervision of Prof. Hongbin Zha . As a student, I have been fortunate to be an intern at UiSee, SenseTime, and NVIDIA.

Email  /  Github  /  Twitter  /  Google Scholar  /  LinkedIn  /  CV

We propose Place Recognition Anywhere Model (PRAM) for efficient large-scale localization which automatically defines landmarks in any scenes and recognizes these landmarks for both coarse and fine localization. Previous works of semantic-aware features SFD2 and geometric-aware matcher IMP are used.

A NeRF-based localization pipeline with sparse rendering for high efficiency. Previous works semantic-aware features SFD2 and geometric-aware matcher IMP are used.

We mebed geometric constraint into graph-based matcher (e.g. SuperGlue) to make it work more accurate and robust in challenging conditions (e.g. large viewpoint changes, repetitve textures). Attention scores are used to remove useless keypoints progressively to achieve higher effciency.

Semantics are very useful for local feature detection and description especially for long-term tasks. However, explicit usage of semantics requires segmentation networks and has severe semantic uncertainties. In this paper, we implicitly embed semantics into detection and description to detect robust keypoints and extract semantically-augmented descriptors.

Our first trial on large-scale localization by recognition. We define global instances on building facades which are discriminative for coarse localization and robust to appearance changes. At test time, we recognize these global instances and use them for city-scale localization.

An end-to-end VO system with tracking, remembering and refining components. It works impressively well in autonomous driving and robotics scrnarios.

Our first trial on large-scale localization by recognition. We define global instances on building facades which are discriminative for coarse localization and robust to appearance changes. At test time, we recognize these global instances and use them for city-scale localization.

An end-to-end localization frameowrk which formulates multi-view inputs as a graph and leverages GNN for multi-view information fusion. It works very well in scenarios where a single-view input leads to errors due to similar structures etc.

An end-to-end VO framework with online adaptation at test time to enhance its ability of working in more general environments.