A Survey: Feature Fusion Method for Object Detection Field

Feature fusion techniques represent a critical research topic within the field of computer vision, playing an extensive role in downstream tasks that necessitate rich object representations, such as image classification, semantic segmentation, and object detection. The Feature Pyramid Network (FPN) was proposed in 2017 for object detection, emerged as a groundbreaking work in the area of feature fusion methods. Subsequently, a variety of powerful and innovative feature fusion architectures have been successively proposed, bringing significant performance enhancements in object detection tasks. This paper takes a depth look into feature fusion technologies starting from the domain of object detection. Systematically categorizes existing methods into two main classes based on the intrinsic properties of different feature fusion structures: simple topological fusion structures and complex topological fusion structures. The paper conducts analyses and summarizes the mechanisms of these two types of structures, introduced the evaluation dataset and evaluation indicators, accompanied by a collation of open-source codes for mainstream feature fusion architectures. Ultimately, through systematic review, the paper summarizes the challenges faced by feature fusion methodologies and provides an outlook on future development trends in this area.