How to calculate Balanced Accuracy and Fisher Discriminant Ratio?

[uasghar]

Dear Daniel,

I want to calculate Balanced Accuracy (ba) and Fisher Discriminant Ratio (fdr) for a point cloud classified by qCANUPO. ba and fdr for my trained classifier was ~97% and 9.8, respectively. Does that mean these values are same for the classified point cloud too? My personal thought is that they are not.

In the classified point cloud, the points for which the confidence threshold was lower than 0.90 have been included in both "ground" class and "veg"class. So I don't really know how many points have been correctly or incorrectly classified into each class.

I want to get a quantitative measure of the efficiency/performance of the classification, which in publication is given by ba and fdr. How could I measure these two considering that I don't know how many points have been incorrectly classified into each class?

Regards,

Umair.

[daniel]

Well, I'm not an expert (Dimitri is the author ;) but the issue I see is that to compute the 'ba' and 'fdr' values you need to know the 'ground truth'. We can do it in the training part, because we know that the points from the first cloud all belong to 'class 1' and points from the second cloud all belong to 'class 2'. In the general case, you don't have this information... Am I missing something?

[uasghar]

You're right! I don't have the ground truth for the whole point cloud. I only used a part of the whole point cloud to train my classifier.
Could you please tag Dimitri to this post. I haven't seen him on the forum recently.

Regards,

Umair.