On 1/26/22 09:01, William F Pokorny wrote:
> ...So what's going on.

Well, after digging for a chunk of day, the non-fuzzy result comes from 
a "bug fix" made back in 1998 by Nathan Kopp & CEY in the 
Trace::ComputeReflection() function of trace.cpp.

Basically after the calculation of the reflected ray direction using the 
perturbed normal, they added second dot test of that perturbed reflected 
ray with the raw surface normal.

If the perturbed reflected ray was heading in a direction opposite the 
the raw surface normal it triggers some code which does one of two kinds 
of correction...

Where the perturbed ray direction is opposite the perturbed normal 
direction it simply drops back to reflection rays based on the raw 
normal.  This is where we suddenly get non-normal-perturbed 
reflections...

Otherwise, it pulls the perturbed reflected ray more into alignment with 
the raw surface normal by an amount based upon a negative weighting 
factor and the magnitude of the dot product of the perturb ray direction 
with the raw normal to some degree being stronger up to a cut off set by 
the initial test with the perturbed normal(a).

After all the fix up the reflected ray direction is normalized - which 
is the expected ray direction state.

Ah, what to do... Thoughts anyone? I'm going to have to think about this.

I don't like that we are basically ignoring perturbed results to some 
degree or another - for reflections - beyond certain normal 
perturbations point.

With very rough surfaces - think pile of stones - a reasonable 
expectation would be to get some degree of all internal reflections. A 
darkening due reflections bouncing until they die off inside the surface 
structure. This feels like a more correct result to me.

Guess I need to check whether this bug fix is getting done with 
refraction too...

Bill P.

(a) - Yep, this second fix/pull toward the raw surface normal might not 
always set in a non-opposing direction with the raw surface normal.

(a) Yes, I think this is going to also create some extra 
fuzz/inaccuracies for fresnel effects at glancing angles at near 
tangents to a surface. There, with certain rougher surfaces, we should 
see around the complete 'larger-overall' tangent some.

Post a reply to this message