@rgommers

If you have time / interest check this too, I believe it is a clear improvement and safe to merge.

But let's wait for the people from the issues to comment.

That looks great! I was translating this notebook from scipy to PyTorch and the original Rodrigues vector to unit quaternion implementation introduced some numeric instabiliity that took me a while to hunt down. Your change will solve this issue.

The code that introduced numeric instability:

theta = np.linalg.norm(camera_params[:, :3], axis=1)[:, np.newaxis]
with np.errstate(invalid='ignore'):
    v = camera_params[:, :3] / theta
    v = np.nan_to_num(v)