Feature Rescaling and Fusion for Tiny Object Detection

Jingwei Liu,Xueqi Cheng,Jiafeng Guo,Zhibin Zhang,Yi Gu,Shumin Han

doi:10.1109/access.2021.3074790

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.

Highlights

INTRODUCTIONWith the large development of convolutional neural networks (CNNs) [1], [2], person (pedestrian) detection [3]–[6], [7] is widely and successfully implemented in various computer vision tasks, such as image classification, object detection and pedestrian detection [5]
With the large development of convolutional neural networks (CNNs) [1], [2], person detection [3]–[6], [7] is widely and successfully implemented in various computer vision tasks, such as image classification, object detection and pedestrian detection [5].different from general object detection tasks, which supply a few of clear and distinct objects in an large-scale image, tiny object detection rise more challenging due to the extreme small size of the targets and the large area of the massive background
In tiny person dataset [43], as shown in Fig. 2(a), only rocks are shown in a large area, which are to be recognized as small objects; in Fig. 2(b), a few people are hidden in a ship with a dark background; Fig. 2(c) shows a very small person in a large sea area; in Fig. 2(d), there are two tiny persons are located in the middle of the image

Summary

INTRODUCTION

With the large development of convolutional neural networks (CNNs) [1], [2], person (pedestrian) detection [3]–[6], [7] is widely and successfully implemented in various computer vision tasks, such as image classification, object detection and pedestrian detection [5]. ResNet-50 [41] is used as the backbone, the input images are rescaled into 224 × 224, the scale of the feature map of the layers are decreasing rapidly and the information missing is obviously This is the reason tiny objects have a lower accuracy compared to common object detection such as Ms COCO and VOC. Different from the works that perform better on common size objects, [47] proposes some specific methods focus on small object detection in four ways, which consists of super resolution, regional proposal, representing the targets in different scales and considering the contextual information. An effective approach named SFRF is designed to focus on improving the tiny object detection performance in three aspects.

PREVIOUS WORK

NONPARAMETRIC ADAPTIVE DENSE PERCEIVING

DATASET CHALLENGE

DATA AUGMENTATION

Findings

CONCLUSION