shadow casted by mesh with skeleton not proper

[GameMonetize]

We change the shadows to get far more precision and now we can do self shadowing

but now the shadows are really sensitive to how bones change the world matrix.

 

Sorry I must be stupid but I cannot find the link to the playground

[BitOfGold]

The model is too big for the playground, here is the text file to copy:
 

pcode

[BitOfGold]

Yes, shadows are really nice, I just realized .frustumEdgeFalloff exists, the only problem is with skeleton models.
 

[GameMonetize]

Ok after analysing your bones they seem to produce matrices with a negative determinant.

 

Not a big deal, just set shadowGenerator.bias = -1;

[BitOfGold]

Thanks for helping and checking!

So the exporter is producing wrong matrices?
Or something changed while importing the skeleton?

I made a minimal blender file with skeleton animation and exported it. (attached)
I think there is nothing wrong with it. It cannot be simpler.

Here is the playground:
http://www.babylonjs-playground.com/#DCYRZX#5

So, I can set the bias to -1 and the shadows are working in the playground, but it's not a solution in my game.
I cannot set the bias to -1 everywhere, it just not works, everything is dark then (even the lit sides of the walls), I have to use a value like 0.00002.

I'm stuck with this
 

wiggle.blend

[GameMonetize]

There is something wrong in the matrices because if I manually fix them it works: (see line#31)

https://www.babylonjs-playground.com/#DCYRZX#6

 

Here is the code used to try to clean the indices:https://github.com/BabylonJS/Babylon.js/blob/master/src/Mesh/babylon.geometry.ts#L1003

 

[BitOfGold]

You were right, the wrong matrix weights caused the shadow problems. If sum of weights is not 1.0, the influencer matrix in the end will be bad.

I have re-written this cleaning function, it just distorted the weights. (I think, i really tried to understand...)

This new code normalizes the weights so the sum will be 1.0, no matter if it was less then 1.0 or more.
The result looks good with all meshes I tried - the only thing left is, part of my goblin mesh has zero on all weights. so no influence I think. That part of the mesh still looks bad and inverse, and I don't know how to correct it.

 

        private static _CleanMatricesWeights(matricesWeights: number[]): void {
            if (!SceneLoader.CleanBoneMatrixWeights) {
                return;
            }
            let size = matricesWeights.length;
            for (var i = 0; i < size; i += 4) {
                let weight = 0;
                for (var j = 0; j < 4; j++) {
                    weight += matricesWeights[i + j];
                }
                if (weight>0.0) {
                    let mweight = 1.0/weight;
                    for (var j = 0; j < 4; j++) {
                        matricesWeights[i + j] *= mweight;
                    }
                }
            }
        }

 

[GameMonetize]

The real problem is the one you are mentioning. When all weights are 0. The resulting code will obviously be wrong.

 

Here is an idea:

The bone matrices will contain an unchanged matrix at index bones.length. So perhaps the idea could be to also update matrices indices and just point to this identity matrix.

Does it make sense?

[BitOfGold]

Yes that would be perfect.
Zero influence is "legit" in meshes In my opinion, if I make a flag with a flagpole and weight-paint only the flag to move etc.

One more thing when for example the only weight (of only one bone) is low like 0.1
That means only a little infuence.
Maybe then the weights sould be
0.1 the minimally influencing bone
0.9 the "identity bone"

it can happen with weight paint like my model, so maybe my normalization of weight (if sum is under 1.0) is not correct!
 

 

[GameMonetize]

Correct. Do you want to try to update the function that cleans the matrices?

[BitOfGold]

Now it's working perfectly!
With every mesh I tried.
I had to move the function into babylon.sceneLoader.ts, because the skeletons are not ready when parsing the geometry.

Here is the diff:
https://github.com/BitOfGold/Babylon.js/commit/e0e8a24e9bbdb0d8c4528441f4b82b8b2bee03e7

I don't know how to send a PR from this because there are many other changes in my fork/master.
I think I cannot do that without making another github user and another fork?
 

[GameMonetize]

This is cool!

Basically you can fork bjs again, apply this change and do the PR

[BitOfGold]

PR sent!
(I could not fork another copy from the same repo from github, maybe with some command line git I don't know. But I'm now using another github account)
 

 

[satguru]

@BitOfGold 

Is your PR in? 

I find that I still need to do "shadowGenerator.bias=-0.3;"

[BitOfGold]

@satguru yes, no problem with my models now.
Maybe it's another, different error.
Negative scale can look similar bad with shadows.

A very similar bad transparent/distorted look was with (when using PBR material), the setting of alphaMode (I had to set it 0)

mo.material.subMaterials[0].specularColor = new C3(0.4,0.4,0.4);

mo.material.subMaterials[0].alphaMode = 0;

mo.material.subMaterials[1].specularColor = new C3(0.12,0.12,0.12);

mo.material.subMaterials[1].alphaMode = 0;