The role of the parameter 'scale' in the training mode and the 'scales' in the testing mode

[11710615]

I still have two questions about the training.
1: what' s the input of the training, fully-sampled HR volume or LR volume? The question stems from the fact that the code has
downsampled the h and w directions according to the parameters 'scale' in the 'dataset.py'.
xy_inds = torch.meshgrid(torch.linspace(0, self.H - 1, self.H // self.scale), torch.linspace(0, self.W - 1, self.W // self.scale))
2: what' s the role of the parameter 'scale' in the training mode and the 'scales' in the testing mode? If the input is the LR volume, the model training is converged based on the existing sparse sampling points. It seems that 'scale' is no used in the training process.

Thanks for your fast reply again.

[NarcissusEx]

1. In training stage, our input is a LR volume downsampled the HR one by the 'scale'. After training, we will evaluate the SR performance.
2. In test stage, we assume that checkpoint is the model trained by setting scale=1, and then render the upsampled views related to the scales.

So, the scale in train/eval denotes the downsampling scale while the scale in test denotes the upsampling one. I apologize for causing you any misunderstanding.

[11710615]

I understand that the framework reads a HR volume and downsamples it to the LR one by the parameter 'scale'. But I wonder where the codes represent the downsampling process. Is it the code bellowed?
xy_inds = torch.meshgrid(torch.linspace(0, self.H - 1, self.H // self.scale), torch.linspace(0, self.W - 1, self.W // self.scale))
It seems that this code downsamples the HR volume on x and y direction without changing on z direction, which is opposite to the setting of 'volumetric MISR' in the paper.
Thank you for your patience in replying to my questions.

[NarcissusEx]

We decouple 3D downsampling into xy plane downsampling and z downsampling. Please see

CuNeRF/src/dataset.py

Line 68 in 9da7b0c

z_ind = index // self.scale * self.scale

.

[11710615]

z_ind = index // self.scale * self.scale
It seems the code ensures index is an integral multiple of self.scale. There is no use on downsampling the z direction.
Meanwhile, the index is in the range(self.LEN) which equal to the self.data.shape[0], the possible reason is that the framework reads the LR volume directly, which has been downsampled along the z direction.

[NarcissusEx]

Huh, maybe you can try the following codes:
scale = 4
img_inds = list(range(21))
print ([i // scale * scale for i in img_inds]),
which can only select the index belong to LR volume.

[11710615]

I misunderstood the code.
Thank you very much for your patient reply.

[byungjur96]

Hi I have similar questions about code.
As far as I understand, in case of 3DSR, the input is HR for metric calculation, and 'scale' downsamples input volume into LR volume, which is a real input into a model to be supersampled.
In case of CT volumetric super resolution which only supersamples in z-axis, how can we manipulate input and scale value?