semseg-master的github链接：semseg

一、数据处理

在制作dataset的时候，label直接读取label = cv2.imread(label_path, cv2.IMREAD_GRAYSCALE) # GRAY 1 channel ndarray with shape H * W

调试代码发现读取的是file:///media/D_4TB/zhouhongjie/1.few-shot%20segmentation/3.CaNet/CaNet-master/dataset/dir/VOCdevkit/VOC2012/SegmentationClassAug/2011_002227.png这张图片：

使用label_class = np.unique(label).tolist()调试得到cv读取到的矩阵值：$[0,6,15,255]$，其中255为白色轮廓，0为黑色，6和15分别为两类。

二、训练

经过dataloader直接打包后，经过一些处理后，将input和对应的label送入训练：output, main_loss, aux_loss = model(input, target)

此处model可以选择和下载预训练权重：

三、输出误差和反向传播

四、mIOU计算

之后经过intersectionAndUnionGPU函数计算交集intersection和并集union、target：intersection, union, target = intersectionAndUnionGPU(output, target, args.classes, args.ignore_label)
classes=21，ignore_label=255

def intersectionAndUnionGPU(output, target, K, ignore_index=255):
    # 'K' classes, output and target sizes are N or N * L or N * H * W, each value in range 0 to K - 1.
    assert (output.dim() in [1, 2, 3])
    assert output.shape == target.shape
    output = output.view(-1)
    target = target.view(-1)
    output[target == ignore_index] = ignore_index
    intersection = output[output == target]   # I-iou
    area_intersection = torch.histc(intersection, bins=K, min=0, max=K-1)
    area_output = torch.histc(output, bins=K, min=0, max=K-1)
    area_target = torch.histc(target, bins=K, min=0, max=K-1)
    area_union = area_output + area_target - area_intersection
    return area_intersection, area_union, area_target

需要注意此处交集和并集的计算：除掉忽略的255之后，交集=output[output == target]，并集=output+target-area_intersection

area_intersection = torch.histc(intersection, bins=K, min=0, max=K-1)