RE: [Openexr-devel] HDRI usage questions...

[Nick Porcino]

Thanks Thomas and Greg. It appears that problem and the solution are now both 
identified! I'll just be off to read "Energy Preserving...", if you'll excuse 
me ;-)

-dp

-----Original Message-----
From: Greg Ward [mailto:address@hidden
Sent: Monday, July 14, 2003 8:07 AM
To: Thomas Driemeyer
Cc: Nick Porcino; address@hidden
Subject: Re: [Openexr-devel] HDRI usage questions...


Holly Rushmeier and I developed a non-linear image filter for exactly 
this problem.  It works by spreading out "outlier" samples into 
neighboring pixels.  See our SIGGRAPH '94 paper, "Energy Preserving 
Non-Linear Filters", which is available on the web from 
<http://radsite.lbl.gov/radiance/papers/sg94.2/energy.html>.  The 
technique is implemented and distributed as part of Radiance 
<http://radsite.lbl.gov/radiance> in the program "pfilt".  It only 
operates on Radiance RGBE (a.k.a. HDR) pictures, unfortunately.

-Greg


Begin forwarded message:

> From: Thomas Driemeyer <address@hidden>
> Date: Sun Jul 13, 2003  11:46:37  PM US/Pacific
> To: Nick Porcino <address@hidden>
> Cc: address@hidden
>
> On Sat, 12 Jul 2003, Nick Porcino wrote:
>> As a result when we merge the images together, we are finding that
>> there are very obvious seams since the camera matrices have been set 
>> up
>> differently to generate the various views.
>
> Well, specularity is computed in object space, but the exact point on
> the surface where it is done depends on the eye ray direction, and 
> hence
> on the camera space transformation. You'll see that especially on very
> small or very thin highlights. That's a normal aliasing effect, but 
> with
> a twist: normally aliasing is handled well by adaptive oversampling, 
> but
> with HDR and pixel energies far out of the 0..1 range a tiny change in
> the sample position can have a huge effect on the value stored in the 
> HDR
> image.
>
> Or in other words: if a subsample hits the super-bright highlight, its
> energy is huge, otherwise it's small. Any pixel with at least one huge
> subsample will be huge. If a camera space change makes a huge subsample
> drift into a previously low-energy neighboring pixel, you see the 
> change.
>
> I don't see an easy solution to that. Perhaps a custom 
> energy-conserving
> despeckle filter, or a wider mental ray sample filter such as clipped
> Lanczos, would help? It would have to operate on samples, not pixels.
> (Hm, I wonder whether irregular sampling grids instead of regular pixel
> grids would make a good feature extension for OpenEXR... mental ray can
> write sample files, but it uses a nonstandard compressed format.)
>
> I assume you are not talking about simple terminator effects at 
> triangle
> edges caused by insufficient tessellation. Those are visible without 
> HDR.
>
> Thomas