A Study on Accuracy Improvement of Small Object Detection Using Deep Neural Networks

The accuracy of visual object detection, which es-timates locations and classes of target objects in input images, has been drastically improved by the rapid advancements of deep convolutional neural networks (CNNs). The evaluation of existing methods based on CNNs is usually conducted using major datasets such as MS-COCO and PASCAL-VOC, and these datasets include several sizes of target objects. The accuracy of detecting of larger objects has become remarkable via recent methods; however, it remains difficult for recent CNNs to accurately detect small objects. To address this problem, this study investigates how to improve the accuracy of small object detection using CNNs. For the investigation, two types of datasets that solely comprise small target objects were created: the Bird and SAVMAP datasets that solely include flying objects in the sky and mammals in the savannah, respectively. Experimental results obtained with the datasets indicate that the input size, depth of CNN layers, and surrounding context of target objects were important factors for small object detection. Furthermore, these results demonstrate that EfficientDet-DO achieved accuracies of 0.6585 and 0.6501 for the Bird and SAVMAP datasets, respectively.