An Ensemble Method of CNN Models for Object Detection

Abstract

Object detection is a research field that deals with detecting objects of a certain class in digital images and videos. Traditional methods of object detection were based on pre-structured features and had limitation on accuracy and computational efficiency. As deep learning had been proved to be a breakthrough, researches about object detection method based on deep learning, especially CNN, started. CNN-based object detection methods can be divided into two types. One is two-stage detector that once region proposals are generated, then they are classified. The other is one-stage detector that detects and classifies the object without generating region proposals. In two-stage detector case, combining CNN models is one of the ways to improve the accuracy in detection, which is called ensemble. In ensemble method, the region proposals generated from each CNN models are combined, classified, and finally voted. When selecting CNN models to be used in ensemble method, various properties of them should be considered in order to enhance complementary strength. In this paper, we propose advanced ensemble method in object detection with novel methods of model selecting and box voting. It is proved with experiment that the accuracy in object detection increased with our proposed methods. Also, combining the original method and our proposed method is expected to further increase the accuracy in detection and make ensemble model more robust.