Combining Instance Segmentation and Background Subtraction Models for Small Object Detection

Abstract

Object detection is a fundamental problem in computer vision that has been extensively investigated over the past decades. Although deep neural networks (DNNs) improve object detection, they cannot effectively recognize small objects. Detecting small objects remains challenging owing to several factors, such as low resolution and scale variance. These challenges are particularly evident in object detection using surveillance cameras, where small objects typically appear in cluttered environments and at varying distances. In object detection using surveillance cameras, the static nature of the background has not been fully exploited in DNN-based object detection methods, although it can be an important cue for detection. In this study, we propose a simple yet effective method to enhance object detection in areas of small objects not detectable by state-of-the-art DNN-based instance segmentation methods. The proposed method extracts foreground regions using a background subtraction model and classifies them, thereby enabling the identification of small objects. In our experiments, we evaluate two real-world scenarios: detecting a person walking on campus and identifying vehicles in road-surveillance footage. The results show that our method improves the detection of small objects and performs better than baseline methods.