目标检测iou loss

GIOU Loss

1. 解决 iou loss 为1 时不优化的缺陷
2. 增加c项，表示为预测框与标签框最小矩形
3. 缺陷，当两个框相交时，尤其一个框包含另一个的时候，退化为iou loss，收敛会变慢

 

DIOU Loss

1. 直接最小化Anchor和目标框之间的归一化距离以达到更快的收敛速度
2. b、bgt分别是anchor、groundtrueth中心点坐标，p为其欧式距离，c为其包含最小矩形对角线长度
3. 解决GIOU Loss 退化IOU loss的情况

CIOU Loss

1. 回归损失在与目标框有重叠甚至有包含关系时更准确，收敛更快
2. a是平衡系数参数，v为衡量anchor和目标框比例一致性参数

class IouLoss(nn.Module):

    def __init__(self,pred_mode = 'Center',size_sum=True,variances=None,losstype='Giou'):
        super(IouLoss, self).__init__()
        self.size_sum = size_sum
        self.pred_mode = pred_mode
        self.variances = variances
        self.loss = losstype
    def forward(self, loc_p, loc_t,prior_data):
        num = loc_p.shape[0] 
        
        if self.pred_mode == 'Center':
            decoded_boxes = decode(loc_p, prior_data, self.variances)
        else:
            decoded_boxes = loc_p
        if self.loss == 'Iou':
            loss = torch.sum(1.0 - bbox_overlaps_iou(decoded_boxes, loc_t))
        else:
            if self.loss == 'Giou':
                loss = torch.sum(1.0 - bbox_overlaps_giou(decoded_boxes,loc_t))
            else:
                if self.loss == 'Diou':
                    loss = torch.sum(1.0 - bbox_overlaps_diou(decoded_boxes,loc_t))
                else:
                    loss = torch.sum(1.0 - bbox_overlaps_ciou(decoded_boxes, loc_t))            
     
        if self.size_sum:
            loss = loss
        else:
            loss = loss/num
        
        return loss