Improving Object Detection with Convolutional Neural Network via Iterative Mechanism

Abstract

The iterative mechanism is prevalent and widely used in many fields, since iterations of simple functions can make complex behaviors. But this mechanism is often overlooked by the state-of-the-art convolutional neural network (CNN)-based object detection methods. In this paper, we propose to use the iterative mechanism to improve the object detection performance of the CNN algorithms. In order to show the benefits of using the iterative mechanism in object detection from more aspects, the main contributions of our work are two aspects: Firstly, we train an iterative version of Faster RCNN to show the application of the iterative mechanism in improving the localization accuracy; Secondly, we present a prototype CNN model that iteratively searches for objects on a very simple dataset to generate proposals. The thoughtful experiments on object detection benchmark datasets show that the proposed two iterative methods consistently improve the performance of the baseline methods, e.g. in PASCAL VOC2007 test set, our iterative version of Faster RCNN has 0.7115 mAP about 1.5 points higher than the baseline Faster RCNN (0.6959 mAP).