Hoover metrics: some advices?

Dear all,
I am struggling to use the hoover metrics application: it takes ages on my small 2000x2000 image, I can’t have a result. After running on 6 cpus the entire week end: still running monday, and I stopped it.
Does anyone have some advice to compute these metrics? I just want to compare 2 segmentations (UInt16 images of labels).
Many thanks,

Hi Remi,

I think we used it to compare a segmentation with reference data, in which there are only a few components. The hoover metrics require to build some sort of confusion matrix between tested and reference segmentation, so if you have N segments in first and M in latter, cost is O(NxM), which may be a lot if both input are image segmentation.



1 Like

which means that, in practice, HooverCompareSegementation cannot be used to compare segmentations, but to evaluate one segmentation based on (few) ground truth polygons. Therefore the name is misleading for the user.

The name of the application refers to the paper

Hoover et al., “An experimental comparison of range image segmentation algorithms”, IEEE PAMI vol. 18, no. 7, July 1996.

The application implements the comparison algorithm described in the paper. The comparison framework is explained in the paper.

I don’t think this is misleading.


The article certainly describes a method to compare segmentation results, but OTB’s implementation cannot compare segmentation results in practice:

Therefore, the actual application has a misleading name for the user (who is interested on using it with real data, not demo data) and it should rather be named HooverEvaluationSegmentation. And even there with severe limitations because of computational cost (see Very slow HooverCompareSegmentation). I do not know if these limitations are unavoidable because they are intrinsic to the method, or more efficient implementations are possible.