Terrain precision

[radoslavirha]

Hi,

I implemented terrains in my app and everything was fine, until I added point cloud to map. Half of the model is on terrain, another half is floating (point cloud is precise and person who measured it confirmed it's ok). I'm using some government elevation data (highest precision I can get ... but I must convert it => #2193)

I compared my terrain with the same terrain in Esri viewer and it's different sometimes. Again globally it seems to be fine, we found anomaly in specific location (maybe more as we will add more models). I tried to improve my lerc decoding and creating png inspired by terrarium, mapbox, amazon formulas and it's always the same, so I think I'm not deforming data while converting.

It looks different when using delatin and martini tesselator, changing meshMaxError in both directions does not help.

I can't identify where the issue is, but I want to ask if it can be deformed by tesselator somehow? Or it works only for some max precision?

Esri viewer:

Martini in deck.gl:

Delatin in deck.gl:

Same texture as elevation raster:

As you can see martini creates smooth terrain, but the slope on the left side is missing, delatin creates slope correctly but terrain si very deformed. I covered elevation model with the same texture and there is slope on the left side visible (blue and green color).

So what do you think? From my point of view data seems to be correct but I think something is wrong with tesselators? Yes and elevation model should be something around 20cm. Thanks

[kylebarron]

It looks different when using delatin and martini tesselator

delatin and martini are two different algorithms. From mapbox/martini#15 (comment)

Martini:

• Only works on square 2^n+1 x 2^n+1 grids.
• Generates a hierarchy of meshes (pick arbitrary detail after a single run)
• Optimized for meshing speed rather than quality.

Delatin:

• Works on arbitrary raster grids.
• Generates a single mesh for a particular detail.
• Optimized for quality (as little triangles as possible for a given error).

At a high level, Martini creates only right triangles, so the mesh will be bigger.

changing meshMaxError in both directions does not help

The meshMaxError defines when the algorithm stops refinement of triangles. A larger meshMaxError leads to larger triangles that approximate the terrain less accurately.

[kylebarron]

Half of the model is on terrain, another half is floating

In summary, this is expected because you're using two different terrain sources and because there's approximation inherent in the terrain generation. The way to "solve" it would be to "snap" point cloud data onto the terrain mesh.

[radoslavirha]

I don't want to snap point cloud to the terrain. I just realised (when I added point cloud) my terrain is deformed compared to the same terrain in ESRI viewer. On the first image there is ESRI viewer and ESRI datasource (I expect it to be good). As you can see my point cloud would fit to the terrain. Once I used the same terrain in my deck app, there is a huge gap on the left side and I don't know why.

I don't need point cloud sitting on the terrain with cm precision, but few meters? is too much I think. Terrains are measured with some precision, this should be 20 cm if I remember, Point cloud is much more precise. It won't match.

So I'm curious why it's so different with the same elevation data and what can I do with it. :)

[kylebarron]

The default meshMaxError is 10 meters. So it makes sense that if you're expecting 20cm precision that the terrain will look bad. You can set meshMaxError to 0 but then you'll have an inefficient mesh.

loaders.gl/modules/terrain/src/terrain-loader.ts

Line 19 in d6e430d

meshMaxError: 10,