Feature Rescaling and Fusion for Tiny Object Detection

Recent years have witnessed rapid developments on computer vision, however, there are still challenges in detecting tiny objects a large-scale background. The knowledge become sparse and weak due to their size, which makes the difficult be detected with common approaches. In this paper, new network named Specific Characteristics based Feature Rescaling Fusion (SFRF) is designed detect persons broad horizon massive Different from methods general, Nonparametric Adaptive Dense Perceiving Algorithm (NADPA) automatically select generate resized feature map high density distribution of objects. Then, method called Many-For-One strategy used for fusion pyramid (FPN) layers improve representation detection. Finally, an ensemble model hierarchical Coarse-to-fine mechanism proposed further performance. experiments demonstrate that approach achieves obvious performance improvement object detection than existing approaches, our has been awarded as 1st-place first Tiny Object Detection (TOD) challenge.