Jiaotong University Team Publishes Adaptive Image Enhancement Paper at ICCV

A research team from the Innovation Center for Intelligent Connected Electric Vehicles at Shanghai Jiao Tong University, led by Professor Zhang Song’an, has published a paper titled “Adapt Foundational Segmentation Models with Heterogeneous Searching Space” at the prestigious IEEE/CVF International Conference on Computer Vision (ICCV). The first author of the study is Yi Li, a joint master's student from the Puyuan Future Technology Institute and Contemporary Amperex Technology Co., Limited (CATL), specializing in electronic information. Research Background General foundational segmentation models often struggle with unconventional image domains, such as camouflaged objects and medical images. Fine-tuning these models is typically challenging due to the difficulty in preparing sufficient datasets and the time constraints involved. One potential solution is to enhance images before they are processed by the segmentation models, thereby improving segmentation performance without altering the base models. Research Status Current image enhancement techniques primarily rely on rule-based methods, which are effective but limited in their ability to adapt to diverse scenarios. Learning-based methods, while more versatile and capable of enriching the array of enhancement techniques, often predict non-descriptive adjustments (such as depth estimation). Combining these two approaches creates a heterogeneous searching space, which can address the limitations of both methods. Research Results To tackle these issues, the team proposed the “Adapt To Augment” paradigm. This approach replaces traditional rule-based enhancements with optimal augmentation strategies to optimize segmentation performance. The method leverages 32 different enhancement techniques, including 22 rule-based and 10 learning-based methods, to build a robust and versatile heterogeneous searching space. Distillation technology is employed to speed up the preprocessing phase, ensuring practical application of the optimal strategies. The proposed method significantly improves model adaptability across various domains. Its effectiveness has been validated on nine public datasets: NJU2k, VT1k, CAMO, NC4k, COD10k, Kvasir-SEG, BUSI, KoletorSDDV2, and MTSD. These results demonstrate the method’s potential in enhancing segmentation performance in real-world scenarios. Author Information Yi Li is a master's student in the field of computer vision, reinforcement learning, and segmentation domain adaptation at the Puyuan Future Technology Institute, Shanghai Jiao Tong University. Professor Zhang Song’an is a tenured-track assistant professor at the Puyuan Future Technology Institute and a member of the Innovation Center for Intelligent Connected Electric Vehicles. His research focuses on algorithms for autonomous vehicle decision-making systems. Professor Zhang has published over 30 papers in top journals and conferences, including TITS, TIV, CVPR, and ICCV. He earned his bachelor's and master's degrees in vehicle engineering from Tsinghua University in 2013 and 2016, respectively, and his Ph.D. in mechanical engineering from the University of Michigan in 2021, where he was advised by Professor Peng Hui, the Director of Mcity. After graduation, he worked as a researcher at Ford Motor Company's Robotics Research Institute and served as a committee chair for the Ford-university collaborative project in robotics. He joined Shanghai Jiao Tong University in 2023, continuing his work on intelligent vehicle and robot decision control algorithms, meta-reinforcement learning, industrial embodied intelligence, and AI-assisted aerospace engine design.