About Me

I am currently an Assistant Professor at School of Artificial Intelligence and Robotics, Hunan University (HNU). I received my Ph.D. degree in Communication and Information Systems from Wuhan University (WHU), advised by Prof. Jiayi Ma, and spent one year as a visiting scholar at University of Bern under the guidance of Prof. Paolo Favaro. Prior to that, I earned my B.S. degree in Electronic Information Science and Technology from WHU—a place where I spent ten rewarding years of study and research.

I’m always interested in exploring the integration of deep learning with geometry-based pipelines, with a focus on tasks such as multimodal image matching, visual localization, 3D scene perception, and remote sensing. I am always open to academic discussions and collaborations, whether on these topics or on any related areas that spark your interest. :))

📝 Publications

ICML 2024
sym

Sparse-to-dense Multimodal Image Registration via Multi-Task Learning

Kaining Zhang, Jiayi Ma

International Conference on Machine Learning (ICML), 2024

  • We propose a novel matching paradigm to address the drawbacks that current multimodal image matching paradigms are facing. Experiments are conducted on several cross-modal scenarios, including visible-infrared, visible-near infrared, and GoogleEarth.
AAAI 2024
sym

ResMatch: Residual Attention Learning for Feature Matching

Yuxin Deng, Kaining Zhang, Zizhuo Li, Shihua Zhang, Yansheng Li, Jiayi Ma

Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2024

  • We take a deep look at the attention mechanism of SuperGlue and rethink cross-attention and self-attention from the perspective of traditional feature matching and filtering. Our well-designed attention approach leads to significant performance improvements on several benchmarks.
TITS 2022
sym

Loop Closure Detection via Bidirectional Manifold Representation Consensus

Kaining Zhang, Zizhuo Li, Jiayi Ma

IEEE Transactions on Intelligent Transportation Systems (TITS), 2022

  • This is a comprehensive extension of our ICRA 2021. We speed up our algorithm by 63% and build the incremental database via HNSW. Meanwhile, experiments are conducted on more challenging datasets.
JAS 2022
sym

Loop Closure Detection with Reweighting NetVLAD and Local Motion and Structure Consensus

Kaining Zhang, Jiayi Ma, Junjun Jiang

IEEE/CAA Journal of Automatica Sinica (JAS), 2022

  • We propose AttentionNetVLAD to extract more powerful image representation, and address feature matching via information lying on both a 2D space and an intrinsic manifold.
JAS 2022
sym

Loop Closure Detection via Locality Preserving Matching with Global Consensus

Jiayi Ma, Kaining Zhang, Junjun Jiang

IEEE/CAA Journal of Automatica Sinica (JAS), 2022

  • A simple yet surprisingly effective feature matching approach is proposed for scenes within loop closure detection tasks (e.g., scenes with repetitive structures). The algorithm can provide reliable correspondences within only a few milliseconds.
TITS 2021
sym

Appearance-Based Loop Closure Detection via Locality-Driven Accurate Motion Field Learning

Kaining Zhang, Xingyu Jiang, Jiayi Ma

IEEE Transactions on Intelligent Transportation Systems (TITS), 2021

  • This is a more complete version of our IROS 2021. We propose a more reasonable loss to revise the motion field, and decrease the time complexity of the algorithm from O(N^3) to O(N). Meanwhile, experiments are conducted on more challenging datasets.
ICRA 2021
sym

Appearance-based Loop Closure Detection via Bidirectional Manifold Representation Consensus

Kaining Zhang, Zizhuo Li, Jiayi Ma

IEEE International Conference on Robotics and Automation (ICRA), 2021

  • We attempt to address loop closure detection (LCD) from the semantic aspect to the geometric one. Based on this idea, the proposed LCD system can achieve satisfying results.
IROS 2021
sym

Motion Field Consensus with Locality Preservation: A Geometric Confirmation Strategy for Loop Closure Detection

Kaining Zhang, Xingyu Jiang, Xiaoguang Mei, Huabing Zhou, Jiayi Ma

IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2021

  • We address feature matching in the perspective of motion field recovery, and add extra constraints to the motion field to make it more suitable for scenes within loop closure detection.