TPS: affine seeding

[indecisiveuser]

This is a draft PR. I have been encountering errors with the current (identity) seeding of tiles, which can be upwards of 40k pixels away from the actual tile location. Seeding tiles with little split point coverage with the current affine transformations seems to help. Is this the best way of going about the issue? I'd appreciate your feedback!

[StephanPreibisch]

Why do you suggest to use the approximate affines instead of the actual transform here @indecisiveuser?

[StephanPreibisch]

Hi Lila, I am confused about this one (what do you mean by "current (identity) seeding of tiles"?), I added my first comment in the code that seems to the replace the TPS transform with the approximate affine. Let's discuss it tomorrow?

[indecisiveuser]

Hi, to answer your questions:

1. the current "initial guess" for all the corners is in local coordinates for that tile. But we are looking for the global coordinates
2. only the initial guess and fallback is changed—inverseTol is still there.
3. Yes, let's discuss this tomorrow!

[StephanPreibisch]

Mmmh, why? It uses the inverse of the TPS transform here:

multiview-reconstruction/src/main/java/net/preibisch/mvrecon/process/fusion/blk/BlkThinPlateSplineFusion.java

Lines 464 to 476 in a4c8f1a

	public static Interval inverseTransformedBoundingBox(
	final ThinplateSplineTransform transform,
	final Dimensions dimensions )
	{
	// transforms from source img pixels to render coordinates
	final RealTransform invTransform = new WrappedIterativeInvertibleRealTransform<>( transform ).inverse();
	// estimated bounding box of the source img transformed to render coordinates
	return Intervals.smallestContainingInterval(
	invTransform.boundingInterval(
	new FinalInterval( dimensions ),
	IntervalSamplingMethod.CORNERS ) );
	}

[StephanPreibisch]

more tomorrow :)

[StephanPreibisch]

I do not understand the problem we have ... so it is hard to think about what the right fix is.

[indecisiveuser]

No longer needed after #106