Question concerning Point based Regions, Polygons and Outlines

[dietzc]

Hi @tpietzsch,

is there a particular reason why SamplingIterableInterval has to be of type BooleanType? Because if we only sample on the target RandomAccessible, we don't really have to care about the functionality of the Cursor? We can actually sample over any IterableInterval, independent of the type, right?

Probably I'm missing something.

[dietzc]

another question, going in the same direction: The LabelRegionCursor which implements Cursor<BoolType> could (in theory) return Cursor<VoidType>, as the returned BoolType always true anyway (as we only iterate over the true pixels)?

If I understand it correctly, the BoolType was just choosen, because there is nothing like a VoidType and something has to be returned by the Cursor? Again, I'm sure I'm missing something :-)

Why I'm asking all those questions: I'm currently trying to figure out how I would define a Polygon and a Contour interface which nicely fit into the imglib2-roi architecture. I need the Polygonto describe the SmallestEnclosingRectangle for example and the Contour to describe e.g. the ConvexHull of an object. My idea at the moment is: each Contour is a Polygon, with the property that all subsequent points are 4- or 8-connected. You can derive a Contour from a Polygon by interpolating between subsequent points of the Polygon. However, I want also to be able to sample over them. In the case of a Polygon over each pixel in the Polygon and only over the contour-pixels in case of the Contour. Therefore I was asking myself, which interfaces of imglib2-roi I could use to express these ideas. Maybe you have some advice, for example if I should work with Shape or IterableRegion?

Update: Replace the name of Contour with Outline if you want. This would be more Java-ish I think ;-)

[dietzc]

see: https://github.com/imagej/imagej-ops/tree/geometricops/src/main/java/net/imagej/ops/geometric for some ideas. Maybe some of this work should rather be part of imglib2-roi. I would be happy to contribute, but I'm not sure if my implementations are going into the right direction.

[tpietzsch]

@dietzc Actually, I think that having the ROI-related Iterables base on VoidType is a very very very good idea. I feel a bit stupid for not thinking of it myself ;-)

I didn't actually try to put it into the existing code, but I'll make a branch and do that.

For now, I cannot foresee no real downsides of doing this. And even if there are downsides, I think it is worth it: This will help to avoid confusion with calling Regions.iterable() vs. Views.iterable() on a BooleanType RAI, where Regions.iterable() would only iterate the true pixels while Views.iterable() would iterate all pixels in the interval. Now, a method like Regions.sample() will simply require a IterableInterval<VoidType> parameter, so if you used Views.iterable() on a BooleanType RAI and passed it to sample it simply won't compile.

[tpietzsch]

@dietzc Regarding Polygon and Contour: To me both are defined in real coordinates. You said "My idea at the moment is: each Contour is a Polygon, with the property that all subsequent points are 4- or 8-connected." This suggests that somehow you mix in integer coordinates here. I would not do this for Polygon or ConvexHull.

Also: Contour would be defined by tracing the border pixels of an (integer) connected region? Then it's not necessary convex? Anyway, this kind of Contour would be an IterableRegion to me, because the iteration order (tracing along the border) is the important information. The Polygon would be RealRandomAccessible(Interval)<T extends BooleanType<T>>. (You could get a discretized mask using Views.raster())

[dietzc]

thanks for closing