Question about the code:

[Gwencong]

Thank you for your excellent work. I have two question about the code:

1. Why not preprocess the original normal maps in the dataset to generate normal maps at three different resolutions. Instead, the code down-samples the original normal maps to 1/4 resolution in the GoMVS forward function and then up-samples them to each stage. Does this approach potentially increase the interpolation error of the normal maps?

2. I'm interested in the method of depth-to-normal. The paper mentions using depth of the previous stage to calculate normal for the next stage. How is the normal obtained in the first stage? Or is the first stage directly using a standard cost regularization approach?

3. There seems to be a bug in the GoConv3D and GoUpConv3D codes, an error will be reported when the batchsize is 2

I'm looking forward to your response. Thank you!

[liucsg]

I also encountered the same situation about problem 3.