Von mises distribuiton in RefNeRF

[Dharmendra04]

Why did you parameterize the reflection vector as a distribution, rather than using a single vector in Ref-NeRF.

'Next, we enable the directional MLP to reason about materials with different roughnesses by encoding a distribution of reflection vectors instead of a single vector' I saw this paragraph in your article but couldn't be able to get the intuition of why using a distribution to represent a single reflection vector. what is the advantage in it?