so3_near_identity function

[capoe]

Thank you so much for making this code available. Great paper!

One question regarding the so3_near_identity function in so3_utils.py: https://github.com/dmklee/image2sphere/blob/main/src/so3_utils.py#L51

Could you check that the implementation given here is as intended?

For example, alpha's linspace covers the entire [0, 2*pi] interval, which is unexpected given the function's stated purpose. Also, what is your motivation for the gamma adjustment C = preC - A?

Many thanks

[dmklee]

Sorry for the slow response. That function is from an old implementation of Spherical CNN, see here.

The implementation is as intended. I think the confusion comes from how the Euler angles are parametrized (E3NN uses YXY parametrization). With +Y direction as the north pole, think of alpha as the longitude and beta as latitude. The gamma adjustment is not important for Image2Sphere since max_gamma=2*pi in so3_near_identity_grid.