A unified and costless approach for improving small and long-tail object detection in aerial images of traffic scenarios

Zhongxia Xiong,Tao Song,Shan He,Ziying Yao,Xinkai Wu

DOI: https://doi.org/10.1007/s10489-022-04108-9

IF: 5.3

2022-10-27

Applied Intelligence

Abstract:Object detection is an important and challenging task for the utilization of arterial images. However, in traffic scenarios, small-sized and long-tail distributed object detections are still major challenges for practical applications. Most previous studies consider these two important problems as irrelevant issues and only tackle one of them at a time. In addition, most of these works achieve improvements at the cost of significant computation increase. In this article, we unveil that both object size and frequency should be taken into consideration in a unified manner. Then, we intend to simultaneously improve small-sized and long-tail distributed object detection at training stage. In specific, (1) we first propose a scale-aware label assignment strategy to focus on relatively small objects, it intuitively adjusts sampling areas according to the size of target. (2) Based on the assignment results, a dynamic threshold is generated and subsequently integrated into a dynamic class suppression loss, which suppresses unnecessary negative gradient to boost the performance on tail categories from a statistic-free perspective. We conduct extensive experiments on typical datasets, i.e., VisDrone and UAVDT, results show that the proposed method achieves 6.5% and 7.6% improvements over the baseline, which sets a new state of the art. We further test the proposed approach for real-time detection on an edge device and also show effectiveness of the proposed method. Without aggravating time and computation burden, our method effectively enhances the performance of object detection in aerial images of traffic scenarios.

computer science, artificial intelligence