Yongchao Xu | 徐永超

I am currently a Ph.D. student (2025-2028) majoring in Electronic and Information Engineering at University of Science and Technology of China (USTC), supervised by Prof. Zheng-Jun Zha and Associate Prof. Jiawei Liu. Before that, I pursued my Master’s degree at University of Science and Technology of China (USTC) (2023-2025) and received my Bachelor's Degree from Northeastern University, majoring in Automation (2019-2023).

My research interests include Human-Object Interaction (HOI) detection, Multimodal Large Language Models, and Embodied Perception.

Email: yongchaoxu@mail.ustc.edu.cn

Google Scholar  /  Github  /  Email

We propose a semantic-diversified and interaction-focused framework for open-vocabulary HOI detection, which enhances generalization to unseen interactions while capturing more discriminative and spatially interpretable interaction cues.

We propose a prompt-guided self-training sampling policy for active prompt learning, which improves sample selection by integrating soft actor-critic optimization with real-pseudo hybrid rewards and vectorized critics.

We propose a comprehensive context learning framework for zero-shot HOI detection, which improves recognition by collaboratively modeling semantic and spatial interactions.

We propose a dual-curriculum active prompt learning framework that improves adaptation by progressively selecting informative samples and leveraging reliable pseudo-labeled data.

We introduce a novel HOI detection framework with a well-designed Mamba-based decoder to mine the benefts of existing methods and enhance the ability to recognize diffcult HOI samples.

We propose the frst method to address the issue of multi-task test-time adaptation for pre-trained VLMs. It can also seamlessly migrate to basic single-task scenarios.

We establish the first large-scale, multi-domain Chinese multi-modal fact-checking dataset, encompassing all types of misinfomation in the multi-modal fact-checking task.

The template of this page is from Jon Barron.