Improving the Accuracy of Object Detection Models using Patch Splitting

Abstract

The classification and afterwards also the localization of objects in images were among the initial breakthroughs that stimulated research into artificial neural networks. Improving the detection accuracy along with other metrics has been a continuous research objective and analysis subject. One particular issue during the training of such models is the varying resolution of input images and it's effects on the training/detection results as most models accept arbitrary resolutions and resize them internally to a specific resolution. This article examines the benefits of automatically splitting images into patches before the detection stage and afterwards joining the patches along with the detection results into the original image. An accuracy test is conducted in a real environment.