Experimental Study on the Effect of Loss Function on Object Detection

Abstract. As a classic task in the field of computer vision, the purpose of Object Detection is to find out all the objects of interest in the image and determine their location and size. With the development of neural network technology, Object Detection has entered the era of deep learning. At present, the object detector is divided into two categories: Two-Step detector and One-step detector. The loss function is an important part of the object detector. This paper introduces six loss functions: Smooth L1 Loss, Balanced L1 Loss, IoU Loss, GIoU Loss, DIoU Loss and CIoU Loss. Among these six loss functions, we select the three loss functions of Smooth L1 Loss, Balanced L1 Loss and IoU Loss to experiment. The main purpose of this experiment is to explore and compare the effect of loss function on the performance of Object Detection algorithm. In the experiment, Faster-RCNN and Retinanet represent two different kinds of detectors. We introduce the loss function into the detector in turn and evaluate the performance of the detector using the Pascal VOC0712 dataset. In this experiment, we used the object detection toolbox mmdetection. Meanwhile, the evaluation metric used to evaluate the detector are Recall AP and mAP. The experimental results show that some loss functions have the opposite effect on two different detectors.