about camera normalize？

[lulongfei]

Halo, thank you for sharing your great work.
I see the camera normalization in your code, the camera radius is set to a fixed 1.5,
https://github.com/3DTopia/LGM/blob/main/core/provider_objaverse.py#L127

However, the camera radius in my dataset is randomly selected in the range of [1.4, 1.6]. I wonder whether I can use your code to normalize cameras for my dataset. If not, what should I change the code for my dataset?
Thank you very much.

[kiui]

@lulongfei-luffy Hi, if for each object all the images have the same radius (although it's random among different objects), you can still use the current normalization. However, if the images for the same object have different radius, it's wrong to normalize to the same 1.5 value. In the latter case, you may consider only scaling the radius of the first view's camera to 1.5 (while other views are scaled by the same factor), and then rotate them.

[Gyeongjin Kang]

@ashawkey you mean rotate the other views w.r.t the first view?

More precisely, scale all view w.r.t the first view scale (the scale value to make the first view cam position to 1.5), then the radius becomes (1.5 = first view radius / scale, xx / scale, xx / scale ...). And make poses relative to first view.

And normalization is not necessary..? right?

Thanks.