Research Advances on Deep Learning Based Small Object Detection Benchmarks
Small object detection is an extremely challenging task in computer vision.It is widely used in remote sensing,intelligent transportation,national defense and military,daily life and other fields.Compared to other visual tasks such as image segmentation,action recognition,object tracking,generic object detection,image classification,video caption and human pose estimation,the research progress of small object detection is relatively slow.We believe that the constraints mainly include two aspects:the intrinsic difficulty of learning small object features and the scarcity of small object detection benchmarks.In particular,the scarcity of small object detection benchmarks can be considered from two aspects:the scarci-ty of small object detection datasets and the difficulty of establishing evaluation metrics for small object detection.To gain a deeper understanding of small object detection,this article conducts a brand-new and thorough investigation on small object detection benchmarks based on deep learning for the first time.The existing 35 small object detection datasets are intro-duced from 7 different application scenarios,such as remote sensing images,traffic sign and traffic light detection,pedestri-an detection,face detection,synthetic aperture radar images and infrared images,daily life and others.Meanwhile,compre-hensively summarize the definition of small objects from both relative scale and absolute scale.For the absolute scale,it mainly includes 3 categories:the width or height of the object bounding box,the product of the width and height of the ob-ject bounding box,and the square root of the area of the object bounding box.The focus is on exploring the evaluation met-rics of small object detection in detail from 3 aspects:based on IoU(Intersection over Union)and its variants,based on aver-age precision and its variants,and other evaluation metrics.In addition,in-depth analysis and comparison of the perfor-mance of some representative small object detection algorithms under typical evaluation metrics are conducted on 6 datas-ets.These categories of typical evaluation metrics can be further subdivided,including the evaluation metric plus the defini-tion of objects,the evaluation metric plus single object category.More concretely,the evaluation metrics plus the definition of objects can be divided into 4 categories:average precision plus the definition of objects,miss rate plus the definition of objects,DoR-AP-SM(Degree of Reduction in Average Precision between Small objects and Medium objects)and DoR-AP-SL(Degree of Reduction in Average Precision between Small objects and Large objects).For the evaluation metrics plus single object category,it mainly includes 2 types:average precision plus single object category,OLRP(Optimal Localiza-tion Recall Precision)plus single object category.These representative small object detection methods mainly include an-chor mechanism,scale-aware and fusion,context information,super-resolution technique and other improvement ideas.Fi-nally,we point out the possible trends in the future from 6 aspects:a new benchmark for small object detection,a unified definition of small objects,a new framework for small object detection,multi-modal small object detection algorithms,rotat-ing small object detection,and high precision and real time small object detection.We hope that this paper could provide a timely and comprehensive review of the research progress of small object detection benchmarks based on deep learning,and inspire relevant researchers to further promote the development of this field.
small object detectiondeep learningevaluation metric of small objectssmall object datasetthe defini-tion of small objectssmall object detection benchmark
国家科技期刊平台
NSTL
万方数据
