[osg-users] Suggestions regarding OSG pssm

Fri May 16 03:01:52 PDT 2008

Please check out and test,

problems, i working on:
flickering, angle depend bug (camera does not cover all correct)

...

NVidia, ATI should work without choosing right plattform

2008/5/15 Adrian Egli OpenSceneGraph (3D) <3dhelp at gmail.com>:

> Hello David,
>
> Great to get you email, in fact the current PSSM implementation is still
> buggy, and i would like to solve the bugs for a long time. but unfortunately
> i don't get the free time slot to solve all the bugs inside the pssm. may we
> can work out all problems. so let's start by discussing your
> comments/questions....
>
> 2008/5/15 David Roger <david.roger at oktal-se.fr>:
>
>> Hello Adrian,
>>
>> Thank you for the nice work on pssm.
>> I have looked a little in your pssm implementation, and I have some
>> suggestions and questions.
>>
>> It would be great if you had a little time to read them, I would like to
>> have your feedback on those.
>> Then, if we agree on some modifications, I may be able to help you with
>> their implementation if you wish.
>>
>> 1) There are problems with transparency in the shader: shadows should not
>> affect it. I propose the following change:
>>
>
>
> This isn't a really problem of PSSM or common SM algorithm, this is more a
> design question based on your database / or how you like to render the
> scene. so i don't like to get this shader directly into pssm, but what we
> could to is to add something like this,
>
> if ( transparent_and_opac_shadow) my code else your code, and the
> transparent_and_opac can be a class / instance variable set before
> generating the pssm glsl shader. so we can append then a flag to osgshadow
> or your own programm to enable or disable the "correct" shadow :-)
>
> => Just to implement, you could append this
>
>
>>
>>
>> replace the line :
>> sstr << "    gl_FragColor = (color*(1.0-enableBaseTextureFilter) +
>> colorTex*enableBaseTextureFilter)*(1.0-0.30*v); "<< std::endl;
>>
>> by:
>> // correct alpha support
>>   sstr << "    color = (color*(1.0-enableBaseTextureFilter) +
>> colorTex*enableBaseTextureFilter); "<< std::endl;
>>   sstr << "    gl_FragColor = vec4(color.xyz*(1.0-0.3*v),color.a); "<<
>> std::endl;
>>
>> 2) there is a serious bug when the camera is oriented in some directions.
>> The shadow map frustums seems to be off, and don't cover all the view
>> frustum.
>>
>> I believe this come from your vector product:
>> top = lightDir ^ camLeft;
>> If you handle the case where lightDir and camLeft are parallel or near
>> parallel, the bug might disappear. However I have not implemented any fix
>> yet.
>>
>> I'have included some screenshots of the bug, and it seems to correspond to
>> that case: you can see a gap between the shadow maps.
>>
>
> => we need a bugfix for that, i tried just a workaround (sqrt(2) : point 6
> ) to avoid this, but there is a bug inside, this is true, because i believe
> we don't need sqrt(2)=1.41... as factor to enlarge the shadow map range. we
> need to work out this bug, and also to fix it (of course)
>
>
>>
>> 3) when several shadow map overlap, you do the sum of the shadows.
>>
>>  sstr << "    float v = clamp(";
>>   for (unsigned int i=0;i<_number_of_splits;i++)    {
>>       sstr << "term"    <<    i;
>>       if ( i+1 < _number_of_splits ){
>>           sstr << "+";
>>       }
>>   }
>>   sstr << ",0.0,1.0);"    << std::endl;
>>
>>
>>
>> I think it would be better to do the average, or select one of them.
>> The best would be picking the closest one (more precise) but I had some
>> bugs at the transition, and my fix is not very elegant.
>> Another option would be picking the maximum value :
>> sstr << "    float v = term0;" << std::endl;
>> for(unsigned int i = 1; i<_number_of_splits;i++)    {
>>     sstr << "    v = max(v,term" << i << " );" << std::endl;
>> }
>>
>
> exactly this is also another problem, we still have in pssm, may first we
> should solve the (2) than we can decide to solve this one, may there will be
> a "trick" doing that automatically with
> set depth value range, but i don't know yet
>
>
>>
>>
>> 4) in my tests, the randomness in the filtering did more bad than good, as
>> it introduced a lot of flickering in the shadow when the camera moves. I
>> think there should be an option to disable the use of the random texure.
>>
>
> flickering come from the openscenegraph based, near - far recalcuation (i
> guess) depends on the view angle and the cull visitor updates each frame the
> near far,so may we can fix this also. ( i don't know yet how :-( )
>
>
>>
>> 5) I am not really sure about this one but when you compute the
>> lightCameraSource, you should update zNear and zFar:
>> if ( zNear <= _split_min_near_dist ){
>>           osg::Vec3 dUpdate = -
>> pssmShadowSplitTexture._lightDirection*(fabs(zNear)+_split_min_near_dist);
>>           pssmShadowSplitTexture._lightCameraSource =
>> pssmShadowSplitTexture._lightCameraSource + dUpdate;
>>       }
>>
>> maybe you should add:
>> zNear += dUpdate;
>> zFar += dUpdate;
>>
>>
>> and then replace :
>> pssmShadowSplitTexture._lightNear = 0.1;//zNear;
>> by
>> pssmShadowSplitTexture._lightNear = zNear;
>>
>> I have not tested this change.
>>
>
> this is just to calculate the zFar value , camera positon outside the
> frustum (splitted) and the center of frustum (splitted) is the look at
> (Center) , may you can test it, and may we get better quality of shadow, may
> not. but important is to do such a calculation otherwise we have a clipping
> issue.
>
>
>>
>> 6) I don't see where the factor 1.41425 comes from at the end. I'd like
>> some explanations on this one if you can.
>> (and btw sqrt(2) = 1.41421356)
>>
>
> see (2)
>
>
> so i suggest that we should work out the issue (2) then an point by other,
> to see how the pssm would be.
>
> have a look at the latest .cpp class i submitted this morning.
>
> adrian
>
>
>
> --
> ********************************************
> Adrian Egli

-- 
********************************************
Adrian Egli
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openscenegraph.org/pipermail/osg-users-openscenegraph.org/attachments/20080516/f135d8c2/attachment-0001.htm>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: ParallelSplitShadowMap
Type: application/octet-stream
Size: 6499 bytes
Desc: not available
URL: <http://lists.openscenegraph.org/pipermail/osg-users-openscenegraph.org/attachments/20080516/f135d8c2/attachment-0001.obj>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: ParallelSplitShadowMap.cpp
Type: text/x-c++src
Size: 41130 bytes
Desc: not available
URL: <http://lists.openscenegraph.org/pipermail/osg-users-openscenegraph.org/attachments/20080516/f135d8c2/attachment-0001.cpp>