损失函数：DIOU loss手写实现

下面是纯diou代码

            '''
              计算两个box的中心点距离d
            '''
            # d = math.sqrt((x2 - x1) ** 2 + (y2 - y1) ** 2)
            d = math.sqrt((pred[:, -1] - target[:, -1]) ** 2 + (pred[:, -2] - target[:, -2]) ** 2)

            # 左边x
            pred_l = pred[:, -1] - pred[:, -1] / 2
            target_l = target[:, -1] - target[:, -1] / 2
            # 上边y
            pred_t = pred[:, -2] - pred[:, -2] / 2
            target_t = target[:, -2] - target[:, -2] / 2
            # 右边x
            pred_r = pred[:, -1] + pred[:, -1] / 2
            target_r = target[:, -1] + target[:, -1] / 2
            # 下边y
            pred_b = pred[:, -2] + pred[:, -2] / 2
            target_b = target[:, -2] + target[:, -2] / 2

            '''
              计算两个box的bound的对角线距离
            '''
            bound_l = torch.min(pred_l, target_l)  # left
            bound_r = torch.max(pred_r, target_r)  # right
            bound_t = torch.min(pred_t, target_t)  # top
            bound_b = torch.max(pred_b, target_b)  # bottom
            c = math.sqrt((bound_r - bound_l) ** 2 + (bound_b - bound_t) ** 2)
            dloss = iou - (d ** 2) / (c ** 2)
            loss = 1 - dloss.clamp(min=-1.0, max=1.0)

第一步 计算两个box的中心点距离d

首先要知道pred和target的输出结果是什么
pred[:,:2]第一个：表示多个图片，第二个:2表示前两个数值，代表矩形框中心点(Y，X)
pred[:,2:]第一个：表示多个图片，第二个2:表示后两个数值，代表矩形框长宽(H，W)
target[:,:2]同理，
d =
 

根据上面的分析来计算左右上下坐标lrtb

 然后计算内部2个矩形的最小外接矩形的对角线长度c

 d是两个预测矩形中心点的距离

 下面接受各种极端情况
A 两个框中心对齐时候，d/c=0，iou可能0-1

 A 两个框相距很远时，d/c=1，iou=0

 所以d/c属于0-1
dloss=iou-d/c属于-1到1
因此设置loss=1-dloss属于0-2

 

展示iou\giou\diou代码，这是YOLOX自带的损失函数，其中dloss是我自己写的
YOLOX是下载自
GitHub - Megvii-BaseDetection/YOLOX: YOLOX is a high-performance anchor-free YOLO, exceeding yolov3~v5 with MegEngine, ONNX, TensorRT, ncnn, and OpenVINO supported. Documentation: https://yolox.readthedocs.io/YOLOX is a high-performance anchor-free YOLO, exceeding yolov3~v5 with MegEngine, ONNX, TensorRT, ncnn, and OpenVINO supported. Documentation: https://yolox.readthedocs.io/ - GitHub - Megvii-BaseDetection/YOLOX: YOLOX is a high-performance anchor-free YOLO, exceeding yolov3~v5 with MegEngine, ONNX, TensorRT, ncnn, and OpenVINO supported. Documentation: https://yolox.readthedocs.io/https://github.com/Megvii-BaseDetection/YOLOX

class IOUloss(nn.Module):
    def __init__(self, reduction="none", loss_type="iou"):
        super(IOUloss, self).__init__()
        self.reduction = reduction
        self.loss_type = loss_type

    def forward(self, pred, target):
        assert pred.shape[0] == target.shape[0]

        pred = pred.view(-1, 4)
        target = target.view(-1, 4)
        tl = torch.max(
            (pred[:, :2] - pred[:, 2:] / 2), (target[:, :2] - target[:, 2:] / 2)
        )
        # pred target都是[H,W,Y,X]
        # (Y,X)-(H,W) 左上角
        br = torch.min(
            (pred[:, :2] + pred[:, 2:] / 2), (target[:, :2] + target[:, 2:] / 2)
        )
        # (X,Y)+(H,W) 右下角


        area_p = torch.prod(pred[:, 2:], 1)  # HxW
        area_g = torch.prod(target[:, 2:], 1)

        en = (tl < br).type(tl.type()).prod(dim=1)
        area_i = torch.prod(br - tl, 1) * en
        area_u = area_p + area_g - area_i
        iou = (area_i) / (area_u + 1e-16)

        if self.loss_type == "iou":
            loss = 1 - iou ** 2
        elif self.loss_type == "giou":
            c_tl = torch.min(
                (pred[:, :2] - pred[:, 2:] / 2), (target[:, :2] - target[:, 2:] / 2)
            )
            c_br = torch.max(
                (pred[:, :2] + pred[:, 2:] / 2), (target[:, :2] + target[:, 2:] / 2)
            )
            area_c = torch.prod(c_br - c_tl, 1)
            giou = iou - (area_c - area_u) / area_c.clamp(1e-16)
            loss = 1 - giou.clamp(min=-1.0, max=1.0)

            # pred[:, :2]  pred[:, 2:]
            # (Y,X)        (H,W)
            # target[:, :2]  target[:, 2:]
            # (Y,X)        (H,W)

        elif self.loss_type == "diou":
            '''
              计算两个box的中心点距离d
            '''
            # d = math.sqrt((x2 - x1) ** 2 + (y2 - y1) ** 2)
            d = math.sqrt((pred[:, -1] - target[:, -1]) ** 2 + (pred[:, -2] - target[:, -2]) ** 2)

            # 左边x
            pred_l = pred[:, -1] - pred[:, -1] / 2
            target_l = target[:, -1] - target[:, -1] / 2
            # 上边y
            pred_t = pred[:, -2] - pred[:, -2] / 2
            target_t = target[:, -2] - target[:, -2] / 2
            # 右边x
            pred_r = pred[:, -1] + pred[:, -1] / 2
            target_r = target[:, -1] + target[:, -1] / 2
            # 下边y
            pred_b = pred[:, -2] + pred[:, -2] / 2
            target_b = target[:, -2] + target[:, -2] / 2

            '''
              计算两个box的bound的对角线距离
            '''
            bound_l = torch.min(pred_l, target_l)  # left
            bound_r = torch.max(pred_r, target_r)  # right
            bound_t = torch.min(pred_t, target_t)  # top
            bound_b = torch.max(pred_b, target_b)  # bottom
            c = math.sqrt((bound_r - bound_l) ** 2 + (bound_b - bound_t) ** 2)
            dloss = iou - (d ** 2) / (c ** 2)
            loss = 1 - dloss.clamp(min=-1.0, max=1.0)

            # Step1
            # def DIoU(a, b):
            # d = a.center_distance(b)
            # c = a.bound_diagonal_distance(b)
            # return IoU(a, b) - (d ** 2) / (c ** 2)

            # Step2-1
            # def center_distance(self, other):
            #    '''
            #    计算两个box的中心点距离
            #    '''
            #    return euclidean_distance(self.center, other.center)

            # Step2-2
            # def euclidean_distance(p1, p2):
            #    '''
            #    计算两个点的欧式距离
            #    '''
            #     x1, y1 = p1
            #    x2, y2 = p2
            #    return math.sqrt((x2 - x1) ** 2 + (y2 - y1) ** 2)

            # Step3
            # def bound_diagonal_distance(self, other):
            #    '''
            #    计算两个box的bound的对角线距离
            #    '''
            #    bound = self.boundof(other)
            #    return euclidean_distance((bound.x, bound.y), (bound.r, bound.b))

            # Step3-2
            # def boundof(self, other):
            #    '''
            #    计算box和other的边缘外包框，使得2个box都在框内的最小矩形
            #    '''
            #    xmin = min(self.x, other.x)
            #    ymin = min(self.y, other.y)
            #    xmax = max(self.r, other.r)
            #    ymax = max(self.b, other.b)
            #    return BBox(xmin, ymin, xmax, ymax)

            # Step3-3
            # def euclidean_distance(p1, p2):
            #    '''
            #    计算两个点的欧式距离
            #    '''
            #     x1, y1 = p1
            #    x2, y2 = p2
            #    return math.sqrt((x2 - x1) ** 2 + (y2 - y1) ** 2)




        if self.reduction == "mean":
            loss = loss.mean()
        elif self.reduction == "sum":
            loss = loss.sum()

        return loss

GitHub - Megvii-BaseDetection/YOLOX: YOLOX is a high-performance anchor-free YOLO, exceeding yolov3~v5 with MegEngine, ONNX, TensorRT, ncnn, and OpenVINO supported. Documentation: https://yolox.readthedocs.io/