dists = dists * tf.linalg.norm(rays_d[..., None, :], axis=-1)

[shuochen365]

Hi,

Why multiply each distance by the norm of the light in its corresponding direction to convert it into a real-world distance?

dists = dists * tf.linalg.norm(rays_d[..., None, :], axis=-1)

Can you explain the principle behind it？

Thanks.

[oOXpycTOo]

There is an explanatory diagram I've plotted. I'm not 100% sure it's a correct explanation, but this is how I understand it:

[city19992]

There is an explanatory diagram I've plotted. I'm not 100% sure it's a correct explanation, but this is how I understand it:

I think it is correct.

[kunkun0w0]

Hi,

I think the norm should multiply the ray_d before it adds with ray_o, which seems more reasonable.

[Hugstar]

dists are relative to the direction vector size. So for example, those 3 cases are exactly same for a given ray:

• ray direction = (1, 2, 3), dist = 5
• ray direction = (2, 4, 6), dist = 2.5
• ray direction = (5, 10, 15), dist = 1

To normalize the dists to a unit-vector length, this normalization is performed

[adaxun]

I think we can refer to the sample point generation process to understand the principle . pts = rays_o[..., None, :] + rays_d[..., None, :] * z_vals[..., :, None] and the input of dists is sub by z_vals 。