3D Shadow Shimmering When Moving Objects

[Ahmed Elhamy 109]

Dears,

I am having a shadow shimmering in my scene. I am using shadow mapping technique. I know that there is a lot of posts over the internet dealing with this subject but My issue is slightly different.

I know that the problem comes from sub-texel issue when moving the camera BUT My scene is using a different technique as the camera and the light sources are stable and the objects inside the scene is moving Relative To Eye concept (RTE). The issue is when implementing cascaded shadow mapping and variance shadow mapping as stated in Directx examples, Every thing goes well except that the shadows are shimmering (Flickering). The shimmering is coming from that the objects are moving not the camera. So when I try to solve the problem with adjusting the sub-texel problem with the camera movement it didn't solve the problem as the camera is stable but the objects are not. 

Any help will be appreciated.

Thanks in advance.

[MJP 20018]

This is expected. The typical "stabilization"  techniques for cascade shadow maps only fix flickering for the case of completely static geometry. If the geometry moves or otherwise animates, you'll get flickering due to changes in how the geometry rasterizes into the depth map. The only way to avoid this this is to apply filtering when sampling the shadow map (or potentially pre-filter, if you're using a technique like VSM that supports pre-filtering).

[turanszkij 485]

I answered this question for you on stack exchange.

I think the easiest way is basically to do soft shadows.

[Ahmed Elhamy 109]

Dears,

There is no solution for adjusting the texels size as we did with the camera movement?

[Aressera 2930]

Maybe you could quantize the positions of objects to texels when rendering the shadow map (but only the shadow map)?

[Ahmed Elhamy 109]

Could you please give me more details?

[turanszkij 485]

8 minutes ago, Aressera said:

Maybe you could quantize the positions of objects to texels when rendering the shadow map (but only the shadow map)?

I think you could only do that for ortho shadow maps, not for perspective or omnidirectional. And it would introduce wrong self shadowing anyway if you only do it for the shadow maps.

[JoeJ 2978]

What works pretty well is temporal filtering, this causes temporal lag (the shadow lags behind a moving object), but simmering is totally eliminated. In fact moving objects have a nicer shadow edge than static objects, because static objects can reveal the individual texels even with VSM, but moving objects blur this away nicely.

I get this as a side effect of realtime GI i'm working on, and i assume it is difficult to achieve it with traditional shadow maps.

The easiest way would be to do it in screen space with temporal aliasing (which is the same idea), but i guess you can't get the filter duration long enough to be effective?

The proper way would be to render the shadows to texture maps and apply the filter there, which boils down to texture (or object) space lighting and is probably too much work for just this purpose alone.

Edit:

Hey, you could apply the temporal filter just to the shadow map itself! E.g. shadowMapTexel = shadowMapTexel * 0.95 + newValue * 0.05.

I assume this works well with VSM and would be easy to test. Let me know how it works if you try it... unpractical self shadows?

 

 

Edited September 29, 2017 by JoeJ

[Ahmed Elhamy 109]

 

18 hours ago, JoeJ said:

Hey, you could apply the temporal filter just to the shadow map itself! E.g. shadowMapTexel = shadowMapTexel * 0.95 + newValue * 0.05.

 

Should I add this code to the pixel shader in the first pass, render to shadow map?

This is the code I use in the shader for rendering the shadow map texture.

VS_OUTPUT VSMain( VS_INPUT Input )
{
    VS_OUTPUT Output;
    
    
    Output.vPosition = mul( Input.vPosition, g_mWorldViewProjection );

    return Output;
}

float2 PSMain (VS_OUTPUT Input) : SV_TARGET 
{
    float2 rt;
    rt.x = Input.vPosition.z;
    rt.y = rt.x * rt.x;
    return rt;
}

This is also the code I use to calculate the percent lit in the scene rendering in the variance shadow mapping.

void CalculateVarianceShadow ( in float4 vShadowTexCoord, in float4 vShadowMapTextureCoordViewSpace, int iCascade, out float fPercentLit , in float depth) 
{
    fPercentLit = 0.0f;
    // This loop could be unrolled, and texture immediate offsets could be used if the kernel size were fixed.
    // This would be a performance improvment.
            
    float2 mapDepth = 0;

    // In orderto pull the derivative out of divergent flow control we calculate the 
    // derivative off of the view space coordinates an then scale the deriviative.
    
    float3 vShadowTexCoordDDX = 
        ddx(vShadowMapTextureCoordViewSpace );
    //float3 vShadowTexCoordDDX = vShadowMapTextureCoordViewSpace;
    vShadowTexCoordDDX *= m_vCascadeScale[iCascade].xyz; 
    float3 vShadowTexCoordDDY = 
        ddy(vShadowMapTextureCoordViewSpace );
    //float3 vShadowTexCoordDDY = vShadowMapTextureCoordViewSpace;
    vShadowTexCoordDDY *= m_vCascadeScale[iCascade].xyz; 
    
    mapDepth += g_txShadow.SampleGrad( g_samShadow, vShadowTexCoord.xyz, 
                                       vShadowTexCoordDDX,
                                       vShadowTexCoordDDY);
    //mapDepth += g_txShadow.Sample(g_samShadow, vShadowTexCoord.xyz);
    // The sample instruction uses gradients for some filters.
                
    float  fAvgZ  = mapDepth.x; // Filtered z
    float  fAvgZ2 = mapDepth.y; // Filtered z-squared
    
    if ( vShadowTexCoord.w <= fAvgZ ) // We put the z value in w so that we can index the texture array with Z.
    {
        fPercentLit = 1;
    }
    else 
    {
        float variance = ( fAvgZ2 ) - ( fAvgZ * fAvgZ );
        //variance       = min( 1.0f, max( 0.0f, variance + 0.00001f ) );
        variance = min(1.0f, max(0.0f, variance + 0.00001f));
    
        float mean     = fAvgZ;
        float d        = vShadowTexCoord.w - mean; // We put the z value in w so that we can index the texture array with Z.
        float p_max    = variance / ( variance + d*d );

        // To combat light-bleeding, experiment with raising p_max to some power
        // (Try values from 0.1 to 100.0, if you like.)
        fPercentLit = pow(p_max, 2);
    }
 

Edited September 30, 2017 by Ahmed Elhamy

[JoeJ 2978]

This implies you do not blur the shadow map? You should, because blurring gives soft shadows and soft shadows are usually enough to make the problem acceptable (although popping of individual shadow texels is still visible).

So i'd do the temporal filter on the blurred results which requires to keep a copy of the old blurred shadow map.