Large-scale tile selection for multi-million tile tilesets

[Kaapp]

Hi,

Hope everyone is doing well! My apologies that it's been a while since the conference where we initially proposed the idea - life and work got very busy 😅

We've worked on a tile selection algorithm, in our case for global terrain rendering, that solves for geometric error in the frustum rather than traversing tiles until an acceptable match is found. It's far faster when the tileset is very large (e.g. > 1 million tiles), we experienced the page hanging for a long time when trying to load through the existing traversal algorithm. We'd love to contribute that back to the community in the hopes it might help someone in a similar position.

This selection algorithm is slightly limited in that we have only tested it for 3d-tiles (but it may be possible to adapt to support I3S in the future), and it requires that the tileset must be an implicit tileset so that geometric error can be calculated for an arbitrary level of detail.

It is currently limited in a few additional ways, but these could be supported in the future:

• The tileset must use the REFINE replacement strategy - it does not yet contain logic to select all parent tiles for an ADD strategy.
• It expects a single subtree for which all tiles are available, and all have content - however as long as the subtree data was not too large, a step could be added to remove tiles from the selection which are not available or do not have content.

Before I raised a PR, I wanted to ask for opinions on some architectural questions:

1. How should this be activated? Should this be a prop set by the user when creating the Tileset3D instance, or should it automatically trigger if and only if the above conditions are met? My personal thought was a user-set prop with a runtime error if the tileset does not match the required constraints.
2. Internally how might this be structured? My personal thought was to subclass TilesetTraverser in modules/tiles/src/tileset and override executeTraversal to perform our alternate algorithm, but I was wondering if there was an alternative preference.

Thanks,
Chris