Determination of normal direction in geometry hit function

[trevordblack]

On an unrelated note - I also agree that the determination of the normal direction should probably happen inside sphere:hit.

Originally posted by @vchizhov in #137 (comment)

[trevordblack]

We have a code philosophy question at hand here. At present we have an implicit determination of inside/outside of geometric primitives within the material scatter function.

1. Does it make sense to make this determination explicit?

• e.g. metal, diffuse, and dielectric have to explicitly account for interiority of rays?

2. Should this interiority check moved into the geometric primitive?

[vchizhov]

As far as I am aware there is no determination in the code - if you start rays on the "inside" of an object it will produce wrong results. As an example imagine enclosing your camera inside a sphere, putting a light source inside - then your render will be black, even though it should not be. I think the only determination happens for dielectrics (since you cannot get away with not doing it). If anything doing the same thing for other materials will be only consistent.

As for the check - technically it could happen anywhere where you have access to the intersection ray direction and the normal, but as you pointed out in the previous thread it makes sense to have it happen inside the primitive. Having it there is useful, since this is the very first time where you produce the normal - so if you return the correct facing normal the moment you compute it, you don't have to be worried and reason about anything down the line. The code is as simple as:
normal = dot(normal, ray.dir) < 0 ? normal : -normal;
For performance reasons the correct dot may be kept in the intersection structure, but such a micro-optimization is probably overkill for an introductory ray tracer.
An added benefit is that for dielectrics, you will always get the correct facing normal so your logic will not differ much from the other materials.

TL;DR: Moving the computation of where the normal is facing to the geometry, saves you from logic leaking outside of where the normal is actually computed. It's easier to reason about, prevents bugs, and makes sense.

[hollasch]

There are arguments on both sides (ha! See what I did there?). Some surfaces may have distinct front/back sides (say, blue on the outside, red on the inside). Others have no "inside" (think of a triangle). Some objects have an implicit inside, with respect to index of refraction (glass inside, whatever else outside).

Does material know about inside vs. outside? It's up to the object to know if it has an inside, right? A sphere has an inside, a torus has an inside, a flat quad may or may not have an inside (it can present itself as an infinitely thin box).

Of course, we've given the reader enough for them to play with variants, and we should just go with simplicity where it doesn't matter.

[vchizhov]

I disagree. It seems we are conflating several uses of the normal here. I am not aware of a reasonable scenario, where you want to return the "incorrect" normal, since you should use the "correct" normal for both shading and scattering (if you are aware of such a scenario please inform me of it). If you want to distinguish between a front and back face, this can be easily achieved by adding a bool indicating the "sidedness" to the intersection structure (which would be filled out once again in the intersection with the geometry routine) without coupling this to the normal. This effectively decouples the material from the geometry (while still allowing one to query front/back face).

I'd argue that keeping the correct normal is both simpler and more consistent, especially when comparing dielectrics to other materials. This is obviously a subjective opinion, I just cannot find good enough arguments for the opposite.

[trevordblack]

I think it's necessary to maintain a notion of inside/outside for geometry. I'm pretty sure we're all in agreement there. The question is if we pass a bool on, or if we always pass the outward normal. Both can be used to determine side. I lean toward passing a bool, since in my own understanding, I hadn't grokked that the `normal` returned was always outward until this issue was presented. I had assumed the normal was on the side of the ray. Either way, we need to explain that in the text.

…

On Mon, Dec 2, 2019, 16:53 Vassillen Chizhov ***@***.***> wrote: I disagree. It seems we are conflating several uses of the normal here. I am not aware of a reasonable scenario, where you want to return the "incorrect" normal, since you should use the "correct" normal for both shading and scattering (if you are aware of such a scenario please inform me of it). If you want to distinguish between a front and back face, this can be easily achieved by adding a bool indicating the "sidedness" to the intersection structure (which would be filled out once again in the intersection with the geometry routine) without coupling this to the normal. This effectively decouples the material from the geometry (while still allowing one to query front/back face). I'd argue that keeping the correct normal is both simpler and more consistent, especially when comparing dielectrics to other materials. This is obviously a subjective opinion, I just cannot find good enough arguments for the opposite. — You are receiving this because you were assigned. Reply to this email directly, view it on GitHub <#270?email_source=notifications&email_token=ACNZVKBH2O3Z3LKQGIAH4ITQWWUXZA5CNFSM4JTIPRQ2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEFXWHBQ#issuecomment-560948102>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACNZVKDI2AT6GIEJOGBNHB3QWWUXZANCNFSM4JTIPRQQ> .

[vchizhov]

The front/back notion will be useful if you want to texture the two sides differently as hollasch mentioned. Considering that this isn't used in the book (and that it can be also achieved by duplicating and scaling the geometry), I am not sure whether adding it is necessary. We need the bool for the ior vs 1/ior choice in the dielectric.

Also as an added benefit the correct normal will be consistent with most (all) rendering equation formulations the readers will find (Kajiya, Veach, Dutre, Lessig).

[hollasch]

I think Vassillen makes a good argument. So, would this work?

• Hit record has new value, Boolean inside property, true if the geometry has a concept of inside and the intersection was from the inside.
• The hit record always returns the normal such that N⋅Ro < 0.

Note that we should avoid saying "same side", because a ray outside a sphere needs the outward normal for the front-face intersection, and the inward normal for the back-face intersection.

Any objections?

[vchizhov]

I would recommend having the bool be something like front_face/front_facing or something along those lines. For instance triangles do not have an "inside" but they can have an orientation defined through the order in which the vertices were given. It's not so much that I want to emphasize the naming, but rather that even if a notion of inside is ill-defined, it should still probably produce a reasonable bool for the two sides - for example if you want to render differently based on the side, as you mentioned previously. It's also consistent with the surfaces that have an inside.

And ultimately you need it for refraction. If you have a triangle mesh that has a dielectric material, you need this bool, even if triangles have no inside, but you just rely on the user to provide correctly oriented meshes with no holes (also makes the refraction code more readable I guess).

Just note that this change will induce a change in the dielectric code.

[hollasch]

front?

[vchizhov]

Whatever would make the most sense for people unfamiliar with graphics, I am not good with naming. Obviously this should be explained in more details in the text, and in comments.