Excellent work!
But I find the backbone, which is used to extract feature map, was implemented by yourselves instead of the pretrained resnet18. I wonder why this is happening, as we know pretrained model usually performs better than un-pretrained one, so whether the Imagenet feature is seriously not suitable for MVS?
Excellent work! But I find the backbone, which is used to extract feature map, was implemented by yourselves instead of the pretrained resnet18. I wonder why this is happening, as we know pretrained model usually performs better than un-pretrained one, so whether the Imagenet feature is seriously not suitable for MVS?