Small Object Detection with Improved CenterNet

Abstract

Detection of small objects is a notorious challenge due to their properties of low resolution and less information. Most existing works aim for general object detection while ignoring the impact of the object size, resulting in high missed detection rate of small objects, making it difficult to apply object detection in real scenarios such as autonomous driving. The implementation of an efficient detection algorithm for small objects is thus worth discussing. In this paper, we propose a small object detection framework with improved CenterNet. Specifically, the proposed framework consists of four parts. The super-resolution (SR) subnetwork enhances the information contained in small objects by generating an SR version of the original image. The future pyramid subnetwork fuses different levels of features thereby retaining semantic information of small objects. The center localization header and the size regression header are combined to detect objects via prediction of object center and object size. In this way, the performance of object detection is improved especially on small objects. We have conducted comprehensive experiments on MS COCO benchmark, and the experimental results have proved the effectiveness of our framework, especially in the detection of small objects better than existing methods.