pose_spherical function in clip_utils.py

[marcelsan]

Hi @codestella and @riven314 ,

Thanks for the work on this repo! I just have a question about the function pose_spherical here. Specifically, it is not clear to me how you derive the C2W transformation from phi and theta. I don't understand what is the spherical coordinates system that you put in the comment of the function. Also what is the coordinate system of your NeRF? Is the OpenGL system?
Finally, what is the reasoning behind the matrix transformation: translation, followed by rot_phi, followed by rot_theta?

Thank you!

[mdda]

One other quick question :

Why do you use rng_inputs[1] in each of the three dimensions?

https://github.com/codestella/putting-nerf-on-a-diet/blob/main/nerf/clip_utils.py#L148

It seems that the code should be more like:

    rng, *rng_inputs = jax.random.split(rng, 1+3)
    radius = random.uniform(rng_inputs[0], minval=bds[0], maxval=bds[1])
    theta = random.uniform(rng_inputs[1], minval=-jnp.pi, maxval=jnp.pi)
    phi = random.uniform(rng_inputs[2], minval=0, maxval=jnp.pi/2)
    return pose_spherical(radius, theta, phi)