Purpose of this repo + where we are at.

[dnomadb]

Purpose:

The main idea of this repo is to capture the tools we need to work w/ nodata:

• Seamfilling / blobbing: Inter/extrapolating pixel values into nodata areas.
• Alpha masking: Given a nodata value(s) create a 4th alpha band that contains transparency information.

  What has happened so far:

• I've ported over seamfilling as-is from pxm, and implemented tests. The tests aren't complete, but we are effectively testing the same case as pxm was. PR here: https://github.com/mapbox/nodata/pull/1
• Started w/ @isiyu and @camillacaros a raster-tester module that will contain our various testing tools. These will be crucial for testing in this repo.

  What needs to happen:

• More options for seamfill: As seamfill morphs into a "pre-fill" (extrapolate out the rgb bands into nodata / masked areas, and save as lossy), we'll need to add options for:
  • Bands to fill
  • Output compression type
• * Integrate rio-mucho:* Interpolating data is computationally expensive, and seamfill uses a windowed approach to reduce memory usage (and to noop on areas that have nodata). If we implement rio-mucho, we'll be able to speed this up a good deal.
• Better unit tests / functionalization for seamfilling: Right now we have no unit tests for seamfilling functions. We need to add more, and functionalize out parts of the process in order to test more effectively.
• Better terminology: Seamfilling when we aren't filling seams, blobbing, let's firm up defs so we aren't cornfused when communicating this stuff.

I've experimented with blobbing out and compositing over "actual" data, and the results look promising - there are only tiny artifacts in extrapolated areas:

I'll pick this back up when I am back, but until then, would love for anyone to start poking at this.

cc @mapbox/satellite @isiyu @yhahn

[dnomadb]

@yhahn an update on where I am at:

Right now, I have it set up so that you can specify the (blobbed|filled|smergled) bands. Check out the results (purposefully did a 10px blob on bands 1, 2, and 3):

When we run this through pxm-composite (w/o ze old seamfilling), we get this:

It's not just the alpha band that is masking this - the data from r, g, and b themselves are 0 now:

If we blob all bands then composite, we see two things:

One is that darker edge artifacts still exist - they've just been moved. The second is that there are weird extrapolation artifacts where we've essentially filled too much.

This big question that are coming up for me in pre-blobbing jpegification is: how do we make sure blobbed pixels don't go "over" good pixels? Also, how are we going to preserve the blobbed data values (under masked areas) with the existing compositing method?

Basically - do we also have to make a custom compositor that will handle this? Or am i misunderstanding the sketch you can in your mind?

[dnomadb]

^ This is all good!