SIOU损失函数公式

SIoU损失函数是一种目标检测中常用的损失函数，用于衡量预测框和真实框之间的相似度。其公式如下：

LossSIoU = 1 - IoU + (Δ + Ω) / 2

其中，IoU表示预测框和真实框的交并比，Δ和Ω是两个超参数，用于平衡IoU和中心点距离的贡献。具体而言，Δ用于控制中心点距离的影响，Ω用于控制IoU的影响。

以下是SIoU损失函数的代码实现：

import torch

def box_iou(box1, box2):
    """
    计算两个框的IoU
    """
    # 计算交集的左上角和右下角坐标
    lt = torch.max(box1[:, :2], box2[:, :2])
    rb = torch.min(box1[:, 2:], box2[:, 2:])

    # 计算交集的宽和高
    wh = (rb - lt).clamp(min=0)

    # 计算交集面积和并集面积
    inter = wh[:, 0] * wh[:, 1]
    area1 = (box1[:, 2] - box1[:, 0]) * (box1[:, 3] - box1[:, 1])
    area2 = (box2[:, 2] - box2[:, 0]) * (box2[:, 3] - box2[:, 1])
    union = area1 + area2 - inter

    # 计算IoU
    iou = inter / union.clamp(min=1e-6)
    return iou

def box_ciou(box1, box2, eps=1e-7):
    """
    计算两个框的CIoU
    """
    # 计算交集的左上角和右下角坐标
    lt = torch.max(box1[:, :2], box2[:, :2])
    rb = torch.min(box1[:,2:], box2[:, 2:])

    # 计算交集的宽和高
    wh = (rb - lt).clamp(min=0)

    # 计算交集面积和并集面积
    inter = wh[:, 0] * wh[:, 1]
    area1 = (box1[:, 2] - box1[:, 0]) * (box1[:, 3] - box1[:, 1])
    area2 = (box2[:, 2] - box2[:, 0]) * (box2[:, 3] - box2[:, 1])
    union = area1 + area2 - inter

    # 计算IoU
    iou = inter / union.clamp(min=1e-6)

    # 计算中心点距离
    center1 = (box1[:, :2] + box1[:, 2:]) / 2
    center2 = (box2[:, :2] + box2[:, 2:]) / 2
    dist = ((center1 - center2) ** 2).sum(dim=1)

    # 计算最小外接矩形的对角线长度
    cwh = (box1[:, 2:] + box1[:, :2] - box2[:, 2:] - box2[:, :2]).clamp(min=0)
    c = cwh[:, 0] ** 2 + cwh[:, 1] ** 2
    d = dist / c.clamp(min=eps)

    # 计算CIoU
    v = 4 / (math.pi ** 2) * torch.pow(torch.atan(box1[:, 2] / box1[:, 3]) - torch.atan(box2[:, 2] / box2[:, 3]), 2)
    with torch.no_grad():
        alpha = v / (1 - iou + v)
    ciou = iou - dist / c.clamp(min=eps) - alpha * v
    return ciou

def box_siou(box1, box2, delta=0.1, omega=0.5):
    """
    计算两个框的SIoU
    """
    # 计算IoU
    iou = box_iou(box1, box2)

    # 计算中心点距离
    center1 = (box1[:, :2] + box1[:, 2:]) / 2
    center2 = (box2[:, :2] + box2[:, 2:]) / 2
    dist = ((center1 - center2) ** 2).sum(dim=1)

    # 计算SIoU
    siou = 1 - iou + (delta + omega) / 2 * (dist / (1 - iou + delta + omega))
    return siou.mean()