What does it mean by 'rendering each ray requires querying a multilayer perceptron hundreds of times'

[cx-tian]

To render a single ray, say we sample N=64 points along the ray. Then we only evaluated 64 times, why hundreds of times?

[MvWouden]

NeRF uses Hierarchical volume sampling, see section 5.2 of the original paper for the description.
Long story short, N locations are sampled and put through the "coarse network".
After this, a more "informed" sampling is undertaken by incorporating the output of the coarse network, which tells us which parts of the volume are relevant.
Therefore, more than N samples are taken.

If naive sampling were used, we would indeed get N points along a ray as you expected.

[cx-tian]

@MvWouden Hi thank you, I got it now.