School of Mathematical Sciences Academic Report [2025] No. 037  

(High-Level University Construction Series Report No. 1059)  

Lecture Title: Cross-Domain Few-Shot Depth Estimation

Speaker: Qin Xiaolin (Research Fellow, Chengdu Institute of Computer Applications, Chinese Academy of Sciences)

Date & Time: 10:00–11:00 AM, May 17, 2025  

Venue: Room 2331, Huiwen Building  

Abstract: Although fully supervised depth estimation methods have achieved high performance, their accuracy relies on enormous training costs. In contrast, few-shot depth estimation (FSDE) enables depth prediction in new scenes with only a small number of labeled samples, offering a promising alternative. However, existing FSDE methods primarily target general scenarios and perform poorly in professional domains due to the scarcity of depth information. To address this gap, we extend FSDE to a new task—cross-domain few-shot depth estimation (CD-FSDE)—aiming to transfer depth knowledge from domains with abundant training labels to low-resource domains. For the CD-FSDE task, we propose the Domain Memory-Aware Guidance Network (DMGNet). Specifically, we design a High-order Polynomial Memory (HPM) operator based on the Scaled Legendre (LegS) measure and polynomials, which captures memory features through dynamically updated differential equations. We further introduce a Memory Interaction (MI) paradigm to establish robust correlations between memory features and local features. To ensure training effectiveness, we develop a Gaussian Guidance Factor Generator comprising a Gaussian Domain-space Filter (GDF) and Multi-head Awareness Attention (MAA) to guide the decoding process. Additionally, a fine-tuning inference strategy is introduced to achieve adaptive alignment of cross-domain feature representations and depth information. To the best of our knowledge, this work represents the first attempt to transfer knowledge from the semantic segmentation domain to the downstream task of depth estimation and achieve cross-domain adaptation through few-shot fine-tuning.

Biography: Qin Xiaolin is the Deputy Chief Engineer and Research Fellow at the Chengdu Institute of Computer Applications, Chinese Academy of Sciences (CAS), a professor at the University of Chinese Academy of Sciences, and a Ph.D. supervisor. He is a recipient of the Tianfu Qingcheng Plan (Sichuan Province’s Leading Talent in Scientific and Technological Innovation), a Provincial Academic and Technical Leader, a recipient of the Provincial Outstanding Young Scholars Fund, an overseas high-level talent, and a CAS Western Young Scholar. His research focuses on automated reasoning, algebraic vision, and industry-specific large models. He has led projects such as the National Natural Science Foundation of China, key projects of the National Key R&D Program, CAS STS Programs, and major Sichuan Province AI science and technology projects. Awards include provincial-level first prizes, the Second Prize of the Sichuan Provincial Science and Technology Progress Award, the CAS President’s Excellence Award, the China Industry-University-Research Collaboration Innovation Award, and the China Technology Market Association Golden Bridge Award. He serves as an expert reviewer for the National Science and Technology Progress Award, and reviewer for more than 10 provinces/cities including the Ministry of Science and Technology, the Ministry of Education, the Ministry of Industry and Information Technology, the National Natural Science Foundation of China, and Sichuan Province.  

All faculty and students are welcome to attend!

Invited by: Li Nan

School of Mathematical Sciences

May 14, 2025