Feature Rescaling and Fusion for Tiny Object Detection

Abstract

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

Rethinking the backbone architecture for tiny object detection

“…They therefore modify FPN to a lowlevel feature pyramid network (LFPN) that starts the top-down structure from a middle layer rather than the high layer. Liu et al [Liu et al, 2021b] devise a feature rescaling and fusion (SFRF) network that selects and generates a new resized feature map with a highdensity distribution of tiny objects through the use of a Nonparametric Adaptive Dense Perceiving Algorithm (NADPA) module.…”

A Tiny Object Detection Approach for Maize Cleaning Operations

“…Limited by the lack of visual feature information caused by fewer pixels, the detection accuracy of tiny objects was relatively low [29,30]. In addition, the information loss during the forward propagation of the networks, the uneven distribution of the sample quantities and the setting of anchor boxes, etc.…”