Adioc loss: An Auxiliary descent IoC loss function

Abstract

Object detection based on deep learning has progressed significantly hitherto. Intersection over Union (IoU) is a wildly adopted evaluation metric in this field, which is also used as a loss function to constrain training. To address the problem of IoU zero gradient under the non-overlap circumstance, most improved loss functions tend to change penalty terms while few loss functions tried to improve IoU itself. In this paper, we proposed an intersection over convex (IoC) algorithm via analysis of IoU series loss functions. IoC can provide a gradient when a bounding box (also called a predicted box) and a ground truth (also called a target box) share no region. This characteristic will accelerate the training phase. In consideration of the comprehensiveness of the loss function, we constructed Auxiliary descent intersection over convex (Adioc) loss. Adioc loss function was tested on a one-stage network named Yolov5 and a two-stage network named Faster-RCNN. For the one-stage network, the results showed that under the same training batch, the accuracy of the Yolov5s network on VOC datasets increased by 0.1% ∼ 0.4%, and the accuracy of the Yolov5s network on COCO datasets increased by 0.1% ∼ 0.4% The relative improvement of Yolov5 m is 0.6% ∼ 0.7% on COCO dataset. For the two-stage network, the improvement of Faster-RCNN on COCO datasets is about 0.3% ∼ 0.5%.