EIoU和Focal-EIoU Loss

1. Thesis

Thesis title: "Focal and Efficient IOU Loss for Accurate Bounding Box Regression"

2. Introduction

Although CIoU Loss considers the overlapping area, center point distance, and aspect ratio of the bounding box regression. But the v in the formula reflects the difference in height and width, not the real difference between height and width and their confidence. Therefore, it sometimes prevents the model from effectively optimizing the similarity. In response to this problem, this paper disassembles the aspect ratio based on CIoU, proposes EIoU Loss, and introduces Focal Loss to focus on high-quality anchor frames.

Article contribution:

1. The loss item of the aspect ratio is split into the difference between the predicted height and width and the height and width of the minimum bounding box, which speeds up the convergence speed and improves the regression accuracy;
2. Focal Loss is introduced to optimize the sample imbalance problem in the bounding box regression task, that is, to reduce the optimization contribution of anchor boxes with a low degree of overlap with the target bounding box to BBox regression, so that the regression process can focus more on high-quality anchor boxes .
3. Extensive experiments are conducted on synthetic and real data. Excellent experimental results verify the superiority of the proposed method. Detailed ablation experiments show the impact of different settings of loss function and parameter values.

3、EIoU Loss

EIoU Loss includes three parts: IoU loss, distance loss, height and width loss (overlapping area, center point example, aspect ratio). The height-width loss directly minimizes the difference in height and width between the predicted object bounding box and the ground-truth bounding box, leading to faster convergence and better localization results.

where wc and hc are the width and height of the smallest bounding rectangle between the predicted bounding box and the ground truth bounding box. p is the Euclidean distance between two points.

4、Focal-EIoU Loss

Considering the problem of unbalanced training samples in the process of predicting target bounding box regression, that is, in an image, the number of high-quality anchor boxes with small regression errors is far less than the number of low-quality anchor boxes with large errors. Poor quality anchor boxes will produce excessively large gradients, affecting the training process. The effect of using EIoU Loss directly is not good, so the author combined Focal Loss to propose Focal-EIoU Loss, starting from the gradient point of view, separating high-quality anchor frames from low-quality anchor frames, the formula is as follows:

The authors found through ablation experiments that the best trade-off can be achieved when γ=0.5.
Note: The Focal in this loss is somewhat different from the traditional Focal Loss. The traditional Focal Loss focuses more on training difficult samples, and the more difficult the samples, the greater the loss. The Focal-EIoU Loss is that the higher the IoU, the greater the sample loss, which is equivalent to the weighting effect and helps to improve the regression accuracy.

5. Effect

Although Focal-EIoU Loss occasionally assigns wrong boxes near large objects, overall, themediumandlarge targetThe performance effect on is better than other loss functionsbetter. but yessmall goal，Focal-EIoU Lossslightly inferiorDue to IoU Loss, low-quality boxes and low-confidence predictions may be ignored or assigned to small objects.