Reconsider Classification Approach

[ggetz]

Feature

Coming out of #11341, we noted that there are several limitations of the current classification implementation, including the need for watertight classification volumes, performance issues, transparency issues, and invert classification artifacts.

I've attempted to capture a running list of these with the Classification label.

@lilleyse Has a few longer-term ideas for an improved approach, including taking advantage of WebGPU features like image load/store to do per-pixel linked lists.

[jiangheng90]

so do cesium team have any progress on this?

[ggetz]

@jiangheng90 There hasn't been any activity on this yet.

If you'd like to add more information about your use case, that may help us prioritize. If you are interesting in contributing, please let us know and we'd be happy to discuss implementation or review a PR. Thanks!

[jiangheng90]

@jiangheng90 There hasn't been any activity on this yet.

If you'd like to add more information about your use case, that may help us prioritize. If you are interesting in contributing, please let us know and we'd be happy to discuss implementation or review a PR. Thanks!

I'm interesting in contributing. recently I will review this issue list first. and see what I can do..

[jiangheng90]

@ggetz I have seen the issue of #9419 today, I found the problem is comes from the oit.

so I think what if I add a pass to split the translucent 3dtiles and translucent normal primitive.

In current version, the translucent normal primitive and 3dtiles is excuted in one pass and u can not tell the different type.

I wonder if new draw order is translucent 3dtiles -> translucent tile classification -> translucent primitive all in oit buffer

then composite in oit. In this way u can fix this issue in a very low cost.

but in a long term. I think the drawcommand design will be more and more complecated when one unexpected issue comes out. I think terrain, 3dtiles, normal primitive drawcommd needs a uniform process for shadow, classification, translucent feature...

[ggetz]

Thanks @jiangheng90! @lilleyse would you be able to comment on the suggested approach here?