A question about evaluation

Hi,

I have something unclear about the computation of evaluation metrics. If my algorithm outputs a tooth instance that does not exist in the groudtruth, which metrics will it drag down?

Thanks

My understanding of the localization accuracy(distance between groundtruth and predicted centroids) is that it is defined asymetrically i.e. wrt to the groundtruth centroid, which means a predicted tooth that does not exist in the groundtruth will not participate in the TLA calculation.

On the other hand, Tooth Identification Accuracy(TIA) will first compares the ground tooth centroid with the nearest occurring predicted tooth, which means it is possible that the predicted tooth non-existent in the ground truth might reduced TIA. (the wrong labelled predicted tooth centroid being closer to the groundtruth than the correct labelled predicted tooth centroid).

Please feel free to add to the discussion.

Thanks for your reply !

I think you are right.

And, what about TSA ?
Will a tooth that doesn't exist contribute a "zero" to the sum of F1-scores before computing the average?