cesium_3d_tiles

go/rocktree-to-3d-tiles-cheat-sheet
 -> pdf is good, for starters:
    https://github.com/CesiumGS/3d-tiles/blob/master/3d-tiles-overview.pdf

    https://github.com/CesiumGS/3d-tiles/blob/afa1fa3815c3e3bb97fe26a5e5665186702743ba/README.md


Q: In section 2 of pdf, it says:
    "The `geometricError` property is used to quantify the visual error that woudl occur if the tileset was not rendered.   When the visual error exceeds a certain threshold, then the tileset and the tiles that it contains are considered for rendering.  Details of how the geometric error is used are given in Section 5, "Geometric Error".

   But then in section 5 it says:
    "Each tileset and each tile has a geometricError property that quantifies
     the error of the simplified geometry compared to the actual
     geometry."
    - can be measured / translated into SSE "screen-space error";
    - when SSE exceeds a certain threshold, the runtime will render
      a higher level of detail.

   These do not sound the same-- they sound complementary.
   That is:
     - the section 2 concept is comparing the tile's quality with rendering nothing.  (or, rendering the parent, without this child? or, rendering the next coarser available thing in a set of LODs?  not clear)
     - the section 5 concept is comparing the tile's quality
       with the real geometry, i.e. an infinitely-good version of it.

PA: I think this may is key, in section 5:
    - For a tileset, the geometric error is used to determine whether
      the root tlle should be rendered.  (that sounds like section 2 concept)
      For a tile, the geometric error is used to determine whether the
      children of the tile should be renedered (that sounds like section
      5 concept).
PA: Ah, I think the blurb of section 2 says "is used" which should
    be interpreted loosely-- that is, it's used along with other things,
    it doesn't give a direct value of the visual error that would occur.
    (except, maybe, at root).

Q: what is the section 5 (tile) geometricError, conceptually?
PA:  well, 5.1 shows the relationship beetwen it and sse:

      sse = (geometricError*screenHeight) / (tileDistance * 2 * tan(fovy/2))

    Oh!  It's actual max physical distance of error between
    the representation and the reality, in some sense.

    So it seems like section 5 gE for a tile
    is related to the section 2 gE's of its children?
    How, exactly?

    Well, what happens when we remove a tile from being rendered?
    *if* that tile was the single worst thing (i.e. load bearing) i.e. max sse,
    then the worst thing becomes either:

    AARGH this is complicated to think about!
    Note that each tile can have several children, and also several LODs?
    How are children and LODs related?  I'm totally confused.


Q: in section 2 "Tilesets and tiles" picture,
   it shows content at the leaves only.  That doesnt' seem right/useful;
   I think there must be content at the internal nodes too?
   Ah, yes, section 7 shows a better picture.
   So, when relative error decides it

   Ah cool!  So, a simple list of LODs
   would be expressed as a parent/child/grandchild,
   each with relative error.  Nice!
   So, the decisions of whether to descend depend on both:
   - would I do better by descending? and is the current quality
     not acceptable?
   - and is the child visible?

   This is subtle!
   TODO: write up the criterion and totally understand the optimized
         rendering loop and how it accesses server!