PSPNet 评价指标

1. pix_accuracy

比较预测值的每个像素和实际值的每个像素，若二者相等，则代表预测正确

总数量为实际值的像素总数，求二者比值，即为准确率

def batch_pix_accuracy(predict, target, labeled):
  pixel_labeled = labeled.sum()
  pixel_correct = ((predict == target) * labeled).sum()
  assert pixel_correct <= pixel_labeled, "Correct area should be smaller than Labeled"
  return pixel_correct.cpu().numpy(), pixel_labeled.cpu().numpy()

 

2. IoU_accuracy

计算相同类别的IoU

相交部分：predict == target

因为是多分类：根据类别形成直方图

area_inter = torch.histc(intersection.float(), bins=num_class, max=num_class, min=1)

相并部分：预测 加 实际 减 相交

因为是多分类：

  area_pred = torch.histc(predict.float(), bins=num_class, max=num_class, min=1)

  area_lab = torch.histc(target.float(), bins=num_class, max=num_class, min=1) 

  area_union = area_pred + area_lab - area_inter

def batch_intersection_union(predict, target, num_class, labeled):
  predict = predict * labeled.long()
  intersection = predict * (predict == target).long()
  area_inter = torch.histc(intersection.float(), bins=num_class, max=num_class, min=1)
  area_pred = torch.histc(predict.float(), bins=num_class, max=num_class, min=1)
  area_lab = torch.histc(target.float(), bins=num_class, max=num_class, min=1)
  area_union = area_pred + area_lab - area_inter
  assert (area_inter <= area_union).all(), "Intersection area should be smaller than Union area"
  return area_inter.cpu().numpy(), area_union.cpu().numpy()

 

3. 更新损失

  def _get_seg_metrics(self):
    pixAcc = 1.0 * self.total_correct / (np.spacing(1) + self.total_label)
    IoU = 1.0 * self.total_inter / (np.spacing(1) + self.total_union)
    mIoU = IoU.mean()
    return {
        "Pixel_Accuracy": np.round(pixAcc, 3),
        "Mean_IoU": np.round(mIoU, 3),
        "Class_IoU": dict(zip(range(self.num_classes), np.round(IoU, 3)))}