possible evaluation bug?

Hi!

Possible bug in eval/eval_utils.lua ?

Certain good detections can be skipped because they weren't tagged with 'ok = 1'. For example:

Let there be detections A,B that are assigned to the same object.
A has score 2 with IoU 0.3 and detection B has score 1 with IoU 0.5.

When running addResult() the first one get ok=1, but the second one get ok=0.

Next, when running evaluate() with min_overlap=0.5, box B will be marked as false positive even though it's a true positive.

What do you think?