DeepLab-v3（86.9 mIOU）

论文地址：https://arxiv.org/pdf/1706.05587.pdf（Rethinking Atrous Convolution for Semantic Image Segmentation）

讲解文章：https://blog.csdn.net/qq_14845119/article/details/102942576

参考项目：https://github.com/fregu856/deeplabv3

一、模型

（一）空洞卷积

同v2版本

（二）Going deeper

（三）ASPP with BN ( batch normalization )

v3版本的ASPP相对于v2有了一些改进。

如上图所示，随着rate的变大，有效的卷积区域变得越来越少。在极端情况下，即rate = feature map size时，空洞卷积核的有效卷积区域只有1。为了解决这一问题，作者对ASPP进行了以下改进：

上图中黄色括号括起的部分就是改进之后的ASPP，对于输入的scores map，分别进行五个平行处理：①1×1卷积；②3×3的rate=6的空洞卷积；③3×3的rate=12的空洞卷积；④3×3的rate=18的空洞卷积；⑤全局平均池化+双线性插值上采样。五个操作的输出的尺寸是相同的，对于这五个输出在通道维度上进行concate；然后再进行1×1的卷积。

下面是得到原图大小1/16的scores map的例子：

class ASPP(nn.Module):
    def __init__(self, num_classes):
        super(ASPP, self).__init__()

        self.conv_1x1_1 = nn.Conv2d(512, 256, kernel_size=1)
        self.bn_conv_1x1_1 = nn.BatchNorm2d(256)

        self.conv_3x3_1 = nn.Conv2d(512, 256, kernel_size=3, stride=1, padding=6, dilation=6)
        self.bn_conv_3x3_1 = nn.BatchNorm2d(256)

        self.conv_3x3_2 = nn.Conv2d(512, 256, kernel_size=3, stride=1, padding=12, dilation=12)
        self.bn_conv_3x3_2 = nn.BatchNorm2d(256)

        self.conv_3x3_3 = nn.Conv2d(512, 256, kernel_size=3, stride=1, padding=18, dilation=18)
        self.bn_conv_3x3_3 = nn.BatchNorm2d(256)

        self.avg_pool = nn.AdaptiveAvgPool2d(1)

        self.conv_1x1_2 = nn.Conv2d(512, 256, kernel_size=1)
        self.bn_conv_1x1_2 = nn.BatchNorm2d(256)

        self.conv_1x1_3 = nn.Conv2d(1280, 256, kernel_size=1) # (1280 = 5*256)
        self.bn_conv_1x1_3 = nn.BatchNorm2d(256)

        self.conv_1x1_4 = nn.Conv2d(256, num_classes, kernel_size=1)

    def forward(self, feature_map):
        # (feature_map has shape (batch_size, 512, h/16, w/16))

        feature_map_h = feature_map.size()[2] # (== h/16)
        feature_map_w = feature_map.size()[3] # (== w/16)

        out_1x1 = F.relu(self.bn_conv_1x1_1(self.conv_1x1_1(feature_map))) # (shape: (batch_size, 256, h/16, w/16))
        out_3x3_1 = F.relu(self.bn_conv_3x3_1(self.conv_3x3_1(feature_map))) # (shape: (batch_size, 256, h/16, w/16))
        out_3x3_2 = F.relu(self.bn_conv_3x3_2(self.conv_3x3_2(feature_map))) # (shape: (batch_size, 256, h/16, w/16))
        out_3x3_3 = F.relu(self.bn_conv_3x3_3(self.conv_3x3_3(feature_map))) # (shape: (batch_size, 256, h/16, w/16))

        out_img = self.avg_pool(feature_map) # (shape: (batch_size, 512, 1, 1))
        out_img = F.relu(self.bn_conv_1x1_2(self.conv_1x1_2(out_img))) # (shape: (batch_size, 256, 1, 1))
        out_img = F.upsample(out_img, size=(feature_map_h, feature_map_w), mode="bilinear") # (shape: (batch_size, 256, h/16, w/16))

        out = torch.cat([out_1x1, out_3x3_1, out_3x3_2, out_3x3_3, out_img], 1) # (shape: (batch_size, 1280, h/16, w/16))
        out = F.relu(self.bn_conv_1x1_3(self.conv_1x1_3(out))) # (shape: (batch_size, 256, h/16, w/16))
        out = self.conv_1x1_4(out) # (shape: (batch_size, num_classes, h/16, w/16))

        return out

经过ASPP之后，再通过上线性插值上采样恢复到原图尺寸，就得到了最终的分割图。在v3中作者没有对scores map进行CRF处理。

总的模型过程比较简单，可以分成下面三步：

		feature_map = self.resnet(x) # (shape: (batch_size, 512, h/16, w/16))
        output = self.aspp(feature_map) # (shape: (batch_size, num_classes, h/16, w/16))
        output = F.upsample(output, size=(h, w), mode="bilinear") # (shape: (batch_size, num_classes, h, w))
        return output

二、实验

最高可以达到86.9mIOU