>>>深度学习Tricks，第一时间送达<<<

论文题目：《Focal and Efficient IOU Loss for Accurate Bounding Box Regression》
论文地址： https://arxiv.org/abs/2101.08158

CIOU Loss虽然考虑了边界框回归的重叠面积、中心点距离、纵横比。但是通过其公式中的v反映的纵横比的差异，而不是宽高分别与其置信度的真实差异，所以有时会阻碍模型有效的优化相似性。针对这一问题，有学者在CIOU的基础上将纵横比拆开，提出了EIOU Loss，并且加入Focal聚焦优质的锚框。

EIOU的惩罚项是在CIOU的惩罚项基础上将纵横比的影响因子拆开分别计算目标框和锚框的长和宽，该损失函数包含三个部分：重叠损失，中心距离损失，宽高损失，前两部分延续CIOU中的方法，但是宽高损失直接使目标盒与锚盒的宽度和高度之差最小，使得收敛速度更快。

EIOU Loss优点：

1）将纵横比的损失项拆分成预测的宽高分别与最小外接框宽高的差值，加速了收敛提高了回归精度。

2）引入了Focal Loss优化了边界框回归任务中的样本不平衡问题，即减少与目标框重叠较少的大量锚框对BBox 回归的优化贡献，使回归过程专注于高质量锚框。

实验中，将YOLOv5中的锚框损失函数替换为EIOU Loss，性能远优于原IOU、DIOU以及CIOU等，测试自身数据集发现涨点明显！近期较忙，代码咨询的小伙伴请私聊~

EIOU Loss代码如下：

from __future__ import absolute_import
from __future__ import division
from __future__ import print_functionimport numpy as npimport paddlefrom ppdet.core.workspace import register, serializable
from ..bbox_utils import bbox_iou__all__ = ['IouLoss', 'GIoULoss', 'EIouLoss']@register
@serializable
class Focal_EIoU_Loss(GIoULoss):"""Distance-IoU Loss, see https://arxiv.org/abs/1911.08287Args:loss_weight (float): giou loss weight, default as 1eps (float): epsilon to avoid divide by zero, default as 1e-10use_complete_iou_loss (bool): whether to use complete iou loss"""def __init__(self, loss_weight=1., eps=1e-10, use_complete_iou_loss=True):super(DIouLoss, self).__init__(loss_weight=loss_weight, eps=eps)self.use_complete_iou_loss = use_complete_iou_lossdef __call__(self, pbox, gbox, iou_weight=1.):x1, y1, x2, y2 = paddle.split(pbox, num_or_sections=4, axis=-1)x1g, y1g, x2g, y2g = paddle.split(gbox, num_or_sections=4, axis=-1)cx = (x1 + x2) / 2cy = (y1 + y2) / 2w = x2 - x1h = y2 - y1cxg = (x1g + x2g) / 2cyg = (y1g + y2g) / 2wg = x2g - x1ghg = y2g - y1gx2 = paddle.maximum(x1, x2)y2 = paddle.maximum(y1, y2)# A and Bxkis1 = paddle.maximum(x1, x1g)ykis1 = paddle.maximum(y1, y1g)xkis2 = paddle.minimum(x2, x2g)ykis2 = paddle.minimum(y2, y2g)# A or Bxc1 = paddle.minimum(x1, x1g)yc1 = paddle.minimum(y1, y1g)xc2 = paddle.maximum(x2, x2g)yc2 = paddle.maximum(y2, y2g)intsctk = (xkis2 - xkis1) * (ykis2 - ykis1)intsctk = intsctk * paddle.greater_than(xkis2, xkis1) * paddle.greater_than(ykis2, ykis1)unionk = (x2 - x1) * (y2 - y1) + (x2g - x1g) * (y2g - y1g) - intsctk + self.epsiouk = intsctk / unionk# DIOU termdist_intersection = (cx - cxg) * (cx - cxg) + (cy - cyg) * (cy - cyg)dist_union = (xc2 - xc1) * (xc2 - xc1) + (yc2 - yc1) * (yc2 - yc1)diou_term = (dist_intersection + self.eps) / (dist_union + self.eps)# EIOU termc2_w = (xc2 - xc1) * (xc2 - xc1) + self.epsc2_h = (yc2 - yc1) * (yc2 - yc1) + self.epsrho2_w = (w - wg) * (w - wg)rho2_h = (h - hg) * (h - hg)eiou_term = (rho2_w / c2_w) + (rho2_h / c2_h)        #Focal-EIOUeiou = paddle.mean((1 - iouk + diou_term + eiou_term) * iou_weight)focal_eiou = iouk**0.5 * eioureturn focal_eiou * self.loss_weight

相比于其他目标框回归损失函数，EIOU Loss考虑了重叠面积，中心点距离、长宽边长真实差，基于CIOU解决了纵横比的模糊定义，并添加Focal Loss解决BBox回归中的样本不平衡问题。

C++ 接口到 Java 接口

解决PackageNotFoundError:XXXX异常

计算机维护方面的知识和技巧,电脑系统维护小常识有哪些四大维护电脑的技巧推荐...

PyScript-在HTML裡添加標籤div

oracle 主键删除表_ORACLE中添加删除主键

【转】0-1分数规划等问题

猪年月度总结-「1」

目标检测算法——YOLOv5/YOLOv7改进|将IOU Loss替换为EIOU Loss

>>>深度学习Tricks，第一时间送达<<<

论文题目：《Focal and Efficient IOU Loss for Accurate Bounding Box Regression》 论文地址： https://arxiv.org/abs/2101.08158

目标检测算法——YOLOv5/YOLOv7改进|将IOU Loss替换为EIOU Loss相关推荐

最新文章

热门文章

论文题目：《Focal and Efficient IOU Loss for Accurate Bounding Box Regression》
论文地址： https://arxiv.org/abs/2101.08158