About Me

My name is Jiuhong Xiao. I am currently pursuing my Ph.D. at the Agile Robotics and Perception Lab (ARPL), Tandon School of Engineering, New York University, under the guidance of Prof. Giuseppe Loianno. My research focuses on addressing the challenges of multi-modal image alignment for the applications on robotic perception and localization systems.

I hold a master’s degree in Computer Science from New York University. In the summer of 2020, I worked as a research assistant under Prof. Alfredo Canziani and Prof. Yann LeCun, focusing on autonomous driving perception and control projects. After graduating with my master’s degree, I joined Amazon as an applied scientist, where I contributed to the development of the Just Walk Out (JWO) technology. I also completed multiple internships at Amazon, working on the Amazon One and Amazon DashCart projects.

Prior to that, I earned my bachelor’s degree in Engineering from the Department of Automation at the University of Science and Technology Beijing. In my final undergraduate year, I served as a research assistant at the Intelligent Biomimetic Design Laboratory (IBDLab), Peking University, working on my undergraduate thesis.

Research

UAV Satellite-Thermal Geo-localization

Multi-modal image alignment is critical for UAV thermal geo-localization, especially in nighttime scenarios where GPS may be unavailable. Our research focuses on aligning onboard thermal imagery with reference satellite maps, leveraging techniques such as image matching, homography-based alignment, and uncertainty estimation. These approaches enable real-time, robust localization across a wide range of challenging environments even trained with limited multi-modal paired data.

Related Research:

• STGL (Image Matching): Project
• STHN (Homography-Based Alignment): Project
• UASTHN (Uncertainty-Aware Alignment): Project

Multi-modal Generative Model

We investigate cross-domain generation to bridge modality gaps, such as synthesizing thermal imagery from RGB inputs and vice versa. By disentangling content and style representations, our models can generate photorealistic, geometrically consistent images across modalities, supporting data augmentation and domain adaptation for downstream localization tasks.

Related Research:

• ThermalGen (Style-Disentagled RGB-Thermal Image Generation): Project

Ground-view Visual Geo-localization

Ground-view geo-localization focuses on matching street-level or ground-perspective images to satellite maps for precise position estimation. Our work develops self-supervised and cross-dataset learning strategies that enhance feature robustness and scalability across varying viewpoints and lighting conditions.

Related Research:

• VG-SSL (Self-supervised Learning): Project
• QAA (Enhanced Aggregation for Multi-Dataset Training): Project

Latest News