Object feedback and feature information retention for small object detection in intelligent transportation scenes

Abstract

Object detection is an important task in intelligent transportation scenes, and the performance of small object detection is directly related to its practical application. Small objects typically have poor detection performance due to reasons such as inadequate feature information and localization challenges. This paper proposes object feedback and feature retention, which can effectively improve the performance of small object detection. Firstly, we provide an in-depth analysis of popular methods in small object detection. In order to address the issue that location loss functions are mainly designed for general object detection, this paper proposes the Small Object Intersection over Union (SOIoU) loss function based on object feedback. This function can be adaptively optimized for small objects according to their size, thereby making the model more focused on small objects. In addition, for the issue of insufficient detail information in the output layer, this paper proposes the Small Object Path Aggregation Network (SOPANet) based on feature information retention, which can effectively enhance the information conditions for detecting small objects. Based on the above, the Object Feedback and Feature Retention-You Only Look Once (OFFR-YOLO) model is obtained in this paper. The proposed method improved the small object detection performance by 2.5 % and 4.8 % on the COCO (Common Objects in Context) and BDD100K (Diverse Driving Dataset for Heterogeneous Multitask Learning) datasets, respectively. The experimental results show that the proposed method can effectively improve the detection performance of small objects.