Check if bbox of intersection is valid

[adriantre]

We experienced that the resulting bounding box of the intersection between hits (files) had minx > maxx. This may be because the VectorDatasets consist of points, and the intersection between these two files may be a singularity.

This line does not check the validity, and this was not uncovered until later in the code where dataset.bounds was called, triggering this check.

What I don't understand is why __post_init__ and the check were not triggered when merging the two bounds in box3 = box1 & box2.

torchgeo/torchgeo/datasets/geo.py

Line 971 in 42e42a9

if box3.area > 0 or box1.area == 0 or box2.area == 0:

[adamjstewart]

Can you provide a minimal reproducible example? If not, I can't really figure out why you're seeing minx > maxx...

[adriantre]

I'll see what I can do.

[adriantre]

I get this info second hand from some students, but it looks like it stems from random_bbox_assignment yielding an empty val_dataset. We tested this on a small dataset, so it will be resolved when more data is introduced. Still weird that the check does not get triggered, but I guess it may no be worth spending too much time on resolving it.

[adamjstewart]

This makes sense, I've seen this too with other splitting functions. Let's keep this issue open, it seems to be a common mistake, and Toblerity/rtree#204 makes it much more difficult to understand what is wrong or how to solve it.