cannot request rectangles consisting of one single point

[paulhfu]

Describe the bug

In Volumina, patchAccessor.getPatchesForRect returns the right patches only if the respective rect intersects at least two tiles. If it is intersecting with one tile only, no patches are returned. Also sometimes getPatchesForRect is called with non-nomralized rects and rects of negative topleft corners. In that case the rect is clipped to the bounds of the scene. The clipping does not handle all possible variations of rectangles correctly.

Expected behavior

There should always be at least one patch-number returned if the respective rect is valid and has at least one point in the given scene.

To Reproduce

call patchAccessor.getPatchesForRect with a rect consisting of a single point. It returns an empty list.

Desktop (please complete the following information):

• OS: Ubuntu 18.04

ilastik ilastik/ilastik@29d6a33
lazyflow ilastik/lazyflow@8c2d40b
volumina 2b6551d

[m-novikov]

Can you provide some examples for invalid clipping of rectangles?

[paulhfu]

This is concerned with single valued rects that are multiples of the tilesize. So for example set the tilesize to 128 and call getPatchesForRect(128, 128, 128, 128).
The patches at the boundary are larger, if the image sizes are no multiples of the tilesize. This is also not handled by this function. For example having an image with (400,400) and a tilesize of 128 requesting any rect within 384 and 399 returns an empty list. E.g. call getPatchesForRect(384, 10, 398, 41)

[paulhfu]

Included a fix here #229

[m-novikov]

Please check #233 I added some tests.
Problem with inclusive ranges in my opinion that it breaks expectation from ranges (where end is not inclusive).
Maybe we should always query [x, y, x + 1, y + 1] for point.

[paulhfu]

cool! #233 fixes it. Exclusive end-points make sense.