Debug low-res tile in example

[kylebarron]

Conclusion:

From below

Ok so this is a bug in pymartini in rescale_positions. Specifically, when you set column_row=False you're using a different transposed matrix for the Martini errors step and for the terrain lookup step!

So basically column_row should always be True so that you're using the same axis ordering for generating the mesh and the terrain lookup, and just add better documentation about whether a .T is needed before passing the tile to pymartini.

Todo:

• Remove column_row option; make it always True
• Add docs about needing to transpose input matrix

---

This tile has a way higher mesh max error than it should.

It should be easy to find this tile's OSM index from the example viewer, then it could be anywhere in:

• Terrarium tile (unlikely) Ruled out here
• Decoding (unlikely)
• Mesh generation (possible)
• Encoding (possible)
• Decoding (unlikely)
• Rendering (unlikely) Ruled out here

[kylebarron]

It's also weird how it's the near corner that happens to be lower res on multiple tiles

[kylebarron]

Ok lets debug this tile, in the central grand canyon: zxy=12/772/1607.terrain. Full Url:

https://us-east-1-lambda.kylebarron.dev/dem/mesh/12/772/1607.terrain?url=terrarium&mesh_max_error=18.82

At first glance, you might think the mesh max error of 18.82 is too high, but even if you reduce it to 5, the near corner is still low-res

When visualized with quantized-mesh-viewer, the corner is still low-res, so we can rule out a rendering issue. (Note that the mesh being upside down is expected since it's designed to be the same coordinates as a png in this case for use with deck.gl)

[kylebarron]

To be blatantly obvious, it's clear that the actual topography of that section of the canyon is detailed, so there should be a more detailed mesh

[kylebarron]

Swapping the Mapbox terrain source with Terrarium tiles, it's clear the Terrarium tiles aren't wrong.

[kylebarron]

Bug looks to be in the rescale_positions step. When I use

rescaled = rescale_positions(
    vertices, tile, bounds=bounds, flip_y=flip_y, column_row=True)

I see the correct mesh, just rotated

When I use

rescaled = rescale_positions(
    vertices, tile, bounds=bounds, flip_y=flip_y, column_row=False)

I see it in the "correct direction" (i.e. looking south (north?), which makes sense since I want it to be the same orientation as the PNG), but with incorrect mesh

[kylebarron]

Ok so this is a bug in pymartini in rescale_positions. Specifically, when you set column_row=False you're using a different transposed matrix for the Martini errors step and for the terrain lookup step!

See how sparse the triangles are in the bottom right, but how there are so many in the flat area on the rim!

[kylebarron]

So basically column_row should always be True so that you're using the same axis ordering for generating the mesh and the terrain lookup, and just add better documentation about whether a .T is needed before passing the tile to pymartini.