ZWZiteng Wang

Ziteng Wang

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Ziteng Wang has a total of 4 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • BT-VLoc: Toward Robust and Accurate Visual Localization With Bidirectional Transformer

    by Jinlei Zhuang, Yuxuan Yuan, Shuaiyan Liu, Ziteng Wang, Weiqiang Ma

    Visual localization, a fundamental task of numerous computer vision applications, is solved by scene coordinates regression (SCoRe) techniques. These methods leverage neural networks to predict scene coordinates, followed by a PnP algorithm for camera pose recovery. However, they face performance degradation due to ambiguous visual patterns. Recent methods attempt to solve this issue by integrating global and local features, yet the full potential of bidirectional interactions between these features remains underexplored. In this article, we introduce a visual localization framework based on the bidirectional transformer (BFormer), which adaptively models and integrates global and local features through bidirectional interaction. By employing self-modulated convolutions for local feature extraction and self-attention mechanisms in a downsampled space for global information modeling, BFormer enables a dynamic bidirectional adaptation process that strengthens the interplay between local and global representations. As a result, BFormer effectively distinguishes similar image patches by leveraging the complementary strengths of low-level features that provide geometric details and high-level features that contain semantic information. Benefited from BFormer, BTVLoc realizes impressive localization performance on all 7-Scenes, LIVL, and Cambridge datasets.

  • Open Access

    GCML: Geometric Correlation Encoding Network With Multi-Scale Local Feature Extraction for Accurate Point Cloud Registration

    by Jinlei Zhuang, Ziteng Wang, Weiqiang Ma

    Point cloud registration, a critical component of various applications, aims to establish correspondences between two point clouds. While detector-free methods exhibit outstanding accuracy, they only encode simple geometric features of point clouds while failing to comprehensively model rich geometric correlations between points. Moreover, they treat point clouds as sequential data and only employ the transformer to perform global message propagation during feature enhancement. This approach inevitably neglects the extraction of multi-scale local features, thereby compromising the descriptive capability of point cloud features. This study introduces GCML, a novel detector-free approach that tackles these issues. For the first problem, GCML develops a geometric correlation encoding module (GCEM) that draws inspiration from the Denavit-Hartenberg (DH) modeling method in robotics to effectively encode the geometric correlations between each pair of points within point clouds. This method substantially captures rich geometric structure information inherent in point clouds, thereby improving registration accuracy. For the second problem, GCML proposes a multi-scale local feature perception module (MLFP) that integrates multi-scale local features of point clouds during the feature enhancement stage. Subsequently, GCML employs the transformer architecture to facilitate global message propagation. This design effectively captures both local and global features of point clouds, hence enhancing the descriptive capability of point cloud features. Extensive experiments demonstrate that GCML achieves superior performance on all ModelNet40, DModelNet40, and KITTI datasets.