Off-the-shelf features for the coarse parametric alignment

[rcg12387]

Hi
You wrote in your paper like this: "We use off-the-shelf deep features (conv4 layer of a ResNet-50 network) to obtain
these correspondences."
And in coarseAlignFeatMatch.py you implemented like this:

resnet_feature_layers = ['conv1','bn1','relu','maxpool','layer1','layer2','layer3']
... ... 
resnet_module_list = [getattr(resNetfeat,l) for l in resnet_feature_layers]
last_layer_idx = resnet_feature_layers.index('layer3')
self.net = torch.nn.Sequential(*resnet_module_list[:last_layer_idx+1])

What do you mean exactly about "conv4 layer of a ResNet-50 network"?

[XiSHEN0220]

Hi,
The output of the Layer3 is the so-called Conv4 feature.
The feature map of the 4th convolutional layer (block): conv1, layer1, layer2, layer3