BIT在第六期论文复现赛中如何进行变化检测？
- 内容介绍
- 文章标签
- 相关推荐
本文共计371个文字，预计阅读时间需要2分钟。
相关专题内容，请直接输入关键词或问题，我将提供简洁的解答。
BIT-CD的结构如下图所示。作者通过一个CNN骨干网络（ResNet）从输入图像对中提取高层语义特征，并且使用空间注意力将每个时间特征图转换成一组紧凑的语义tokens。然后使用一个transformer编码器在两个tokens集中建模上下文，得到了富有上下文的tokens被一个连体transformer解码器重新投影到像素级空间，以增强原始像素级特征。最终，作者从两个细化的特征图计算特征差异图像(FDIs)，然后将它们送到浅层CNN以产生像素级变化预测。
class Backbone(nn.Layer, KaimingInitMixin): def __init__(self, in_ch, out_ch=32, arch='resnet18', pretrained=True, n_stages=5): super(Backbone, self).__init__() expand = 1 strides = (2, 1, 2, 1, 1) if arch == 'resnet18': self.resnet = resnet.resnet18( pretrained=pretrained, strides=strides, norm_layer=get_norm_layer()) elif arch == 'resnet34': self.resnet = resnet.resnet34( pretrained=pretrained, strides=strides, norm_layer=get_norm_layer()) else: raise ValueError self.n_stages = n_stages if self.n_stages == 5: itm_ch = 512 * expand elif self.n_stages == 4: itm_ch = 256 * expand elif self.n_stages == 3: itm_ch = 128 * expand else: raise ValueError self.upsample = nn.Upsample(scale_factor=2) self.conv_out = Conv3x3(itm_ch, out_ch) self._trim_resnet() if in_ch != 3: self.resnet.conv1 = nn.Conv2D( in_ch, 64, kernel_size=7, stride=2, padding=3, bias_attr=False) if not pretrained: self.init_weight() def forward(self, x): y = self.resnet.conv1(x) y = self.resnet.bn1(y) y = self.resnet.relu(y) y = self.resnet.maxpool(y) y = self.resnet.layer1(y) y = self.resnet.layer2(y) y = self.resnet.layer3(y) y = self.resnet.layer4(y) y = self.upsample(y) return self.conv_out(y) def _trim_resnet(self): if self.n_stages > 5: raise ValueError if self.n_stages < 5: self.resnet.layer4 = Identity() if self.n_stages <= 3: self.resnet.layer3 = Identity() self.resnet.avgpool = Identity() self.resnet.fc = Identity()
本文共计371个文字，预计阅读时间需要2分钟。
相关专题内容，请直接输入关键词或问题，我将提供简洁的解答。
BIT-CD的结构如下图所示。作者通过一个CNN骨干网络（ResNet）从输入图像对中提取高层语义特征，并且使用空间注意力将每个时间特征图转换成一组紧凑的语义tokens。然后使用一个transformer编码器在两个tokens集中建模上下文，得到了富有上下文的tokens被一个连体transformer解码器重新投影到像素级空间，以增强原始像素级特征。最终，作者从两个细化的特征图计算特征差异图像(FDIs)，然后将它们送到浅层CNN以产生像素级变化预测。
class Backbone(nn.Layer, KaimingInitMixin): def __init__(self, in_ch, out_ch=32, arch='resnet18', pretrained=True, n_stages=5): super(Backbone, self).__init__() expand = 1 strides = (2, 1, 2, 1, 1) if arch == 'resnet18': self.resnet = resnet.resnet18( pretrained=pretrained, strides=strides, norm_layer=get_norm_layer()) elif arch == 'resnet34': self.resnet = resnet.resnet34( pretrained=pretrained, strides=strides, norm_layer=get_norm_layer()) else: raise ValueError self.n_stages = n_stages if self.n_stages == 5: itm_ch = 512 * expand elif self.n_stages == 4: itm_ch = 256 * expand elif self.n_stages == 3: itm_ch = 128 * expand else: raise ValueError self.upsample = nn.Upsample(scale_factor=2) self.conv_out = Conv3x3(itm_ch, out_ch) self._trim_resnet() if in_ch != 3: self.resnet.conv1 = nn.Conv2D( in_ch, 64, kernel_size=7, stride=2, padding=3, bias_attr=False) if not pretrained: self.init_weight() def forward(self, x): y = self.resnet.conv1(x) y = self.resnet.bn1(y) y = self.resnet.relu(y) y = self.resnet.maxpool(y) y = self.resnet.layer1(y) y = self.resnet.layer2(y) y = self.resnet.layer3(y) y = self.resnet.layer4(y) y = self.upsample(y) return self.conv_out(y) def _trim_resnet(self): if self.n_stages > 5: raise ValueError if self.n_stages < 5: self.resnet.layer4 = Identity() if self.n_stages <= 3: self.resnet.layer3 = Identity() self.resnet.avgpool = Identity() self.resnet.fc = Identity()