Problem with dissapearing instances

[sweco-sekrsv]

I have problems with dissapearing instances when they are still clearly in view and within the bounding box when looking in cesium.js. I have tried multiple combinatinations of single models, multiple models and with and without using --use_i3dm. I have also tried with and without textures but the problems persists. The goal is to have multiple tree models working without dissapearing from view when you are up close.

Example command I use:
i3dm.export.exe -c Host=127.0.0.1;Username=postgres;Password=password1234%;Database=fme_test;Port=5432;sslmode=prefer -t globe.tommelilla_trees_export_i3dm -o C:\work\tomelilla\testoutput_cross_cluster7 -g 700 --geometrycolumn geom --max_features_per_tile 2000 --use_external_model true --use_clustering false --boundingvolume_heights 0,30

Do you know if there is any fix I can do to my gltf models to get it working better? I have attached a simple model I tested with.

Lov_asp_1_cr.zip

[sweco-sekrsv]

I tried using use_gpu_instancing. It seems to works better! However the dataset is now 10 gb instead of 100 mb since the same textures is being added to each glb tileset. Would it be possible to just add the meshes into the generated glb files and keep the textures external when using use_gpu_instancing?

[bertt]

There are some known problems with I3dm instances for example see CesiumGS/cesium#11176 but no fix. Maybe you can add something to the discussion there.

GPU instancing should not have this problem, but there are other issues so it's not (yet) the default. I'll take a look at the external textures.

[sweco-sekrsv]

Yeah, external textures would be great for GPU instancing. I have about 10 trees with each of them having a texture of 1024x512 so it really adds alot to the size of the glb tiles.

I have attached an example of a glb with an external texture. If i3dm.export finds a input glb with external texture it could keep the texture external or perhaps with a new flag which then also can support if the input is a gltf file.

In the zip there is also a typescript which is using glTF-Transform which runs clearNodeParent and clearNodeTransform and then writes a glb with external texture. gltfpack can also write with external textures using the -tr flag.

Thanks, I will look at the thread over at Cesium to see if I can add something to the discussion.

tree_example_ext_texture.zip

[sweco-sekrsv]

With some help from gpt 5.2 I managed to build the attached typescript that can patch the output from i3dm.export. It extracts the textures, keeps the extensions intact and write a new glb that reference external textures. The output seem to work great in cesium.js except for picking which seems broken now after my patch.

Folder size before patch: 14.2 gb
Folder size after patch: 97 mb

So keeping the textures external results in 146 times smaller filesize in my case. So definitely worth looking into supporting directly in i3dm.export

patch_i3dm_ext_textures.zip

[bertt]

External textures is implemented in 2.14.1

When the input model (GLB/GLTF) references external image textures (for example *.png), GPU instancing tiles are written as GLB files that reference those textures externally instead of embedding them in every tile.

On export, textures are written once to:

output\\content\\textures\\<modelName>\\<textureFile>
and each generated tile GLB references them via a relative URI:

textures/<modelName>/<textureFile>
This significantly reduces dataset size when many tiles share the same model + textures.

If the input model has embedded images, they remain embedded in each exported tile GLB.

[sweco-sekrsv]

Works great! Thanks for adding this feature that quick. The picking on the features is messed up. Picking one feature will select multiple features, its not even of the same type. As you can see in this example I have multiple input models. Is that a known bug when using gpu instancing?

[bertt]

the Cesium picking function is limited for gpu instancing, but should work alright if you have only 1 model. Did you try that?

[bertt]

I've released 2.15.0 containing some fixes for gpu instancing, not sure it fixes the behavior you saw with picking but worth a try.

[sweco-sekrsv]

I did not see any change in picking behaviour in 2.15.0 for multi models. But the version info still says 2.14.1+4fac22b85a2ba079da6b49d445fbac29b770fb36 on that release. Not sure if you missed to change the build number?

I still have to test single model picking, but my datasets will always have multiple input models.

[bertt]

it's a known Cesium issue multiple input models picker when using ext_mesh_features CesiumGS/cesium#11683

[bertt]

v2.16.0 is released now