How is the IoU for this competition is calculated exactly?

I have pretty diverse local validation data but the local score still doesn’t correlate much with the Leader Board. Do you calculate it per image and average or you calculate by measuring intersection and union over the whole test set data?

Dear @bonlime,

According to the competition webpage, Jaccard micro is to be used where you calculate metrics globally by counting the total true positives, false negatives and false positives.

More details can be given by Dave @daveluo_gfdrr.

1 Like

Yes, @aghandour is correct that like the sklearn `jaccard_score`

with `average=micro`

, we calculate the global intersection and union values before calculating the score rather than averaging the score over all images.

1 Like

Thanks for clarification!