Gltf node instancing not supported in Cesium.

[vinjn]

KhronosGroup/glTF#276

Dae file uploaded as https://cloud.githubusercontent.com/assets/558657/3093329/694702ee-e5b1-11e3-9c48-138430dd89c6.jpg
You can download the jpg, and rename the extension to 7z and extract it.

The expected rendering result should be like

However the result from Cesium is like

[pjcozzi]

Thanks for the report @vinjn

[vinjn]

@pjcozzi Do you have clues on how to debug this issue? I can look into this problem myself with your instructions.

[pjcozzi]

Thanks for the kind offer. In Model.js, have a look at parseNodes() and createRuntimeNodes().

[vinjn]

I have investigated these methods, however no clues have been found so far.

[vinjn]

https://github.com/AnalyticalGraphicsInc/cesium/blob/master/Source/Scene/Model.js#L1833-L1845

// Node hierarchy is a DAG so a node can have more than one parent so it may already exist
var runtimeNode = runtimeNodes[n.id];
if (runtimeNode.parents.length === 0) {
    if (defined(gltfNode.matrix)) {
        runtimeNode.matrix = Matrix4.fromColumnMajorArray(gltfNode.matrix);
    } else {
        // TRS converted to Cesium types
        axis = Cartesian3.fromArray(gltfNode.rotation, 0, axis);
        runtimeNode.translation = Cartesian3.fromArray(gltfNode.translation);
        runtimeNode.rotation = Quaternion.fromAxisAngle(axis, gltfNode.rotation[3]);
        runtimeNode.scale = Cartesian3.fromArray(gltfNode.scale);
    }
}

It seems the transformation fileds of runtimeNode is only updated when gltfNode is parsed for the first time runtimeNode.parents.length === 0.
@pjcozzi is the logic correct here?

[pjcozzi]

This could be the issue. If a node has more than one parent, we should create a new runtimeNode for it with a different set of transforms leading up to the root.

It could be a small fix, but I don't have the bandwidth to look at it right now. Do you want to experiment?

[vinjn]

I would love to have that challenge. When shall we create new runtimeNode then?
My initial plan is:

• create new runtimeNode in parseNodes()
• update transforms in createRuntimeNodes()

[pjcozzi]

I think that will do it. Essentially, in runtimeNode, we will transform the DAG into a tree, then createRuntimeNodes will just work on a tree (no need to check number of parents because it will always be one).

[pjcozzi]

Thanks again @vinjn. This is a good catch since Cesium was trying to handle this case but was not doing so correctly; it had a separate draw call for each instanced node, but they all used the "last" parent's transform.

I hacked up a partial fix in gltf-instancing but it is ugly and only works for multiple immediate parents, not older ancestors. Instead, I'm pretty sure I am going to expand the glTF DAG into a tree, but this will be some work since I will need to map the original glTF node ids (e.g., for animation and show/hide) to the new nodes. @fabrobinet I imagine this is how your renderer works?

[fabrobinet]

I confirm but it's not a big overhead as you can share the same geometry.

[pjcozzi]

Agreed, I'm not concerned about the overhead, just the time I need to update the code :)

[pjcozzi]

@satishkashyap also reported this bug with a different model: KhronosGroup/glTF#363

[pjcozzi]

Fixing this might also fix #2387.

[slozier]

I'm also running into this issue when trying to convert some models. From the glTF issue it seems like the plan is to fix this at the converter level?

[pjcozzi]

Exactly, this will be fixed in the converter. We decided that the glTF spec will not allow DAGs, so DAGs will be converted to a tree offline. See KhronosGroup/glTF#401

This conversion will be part of CesiumGS/gltf-pipeline#1. Work will begin in the new year.

[lilleyse]

The updated converter that should be deployed soon fixes this model.