CODEWITHSUNDEEP
performanceroc
Andreas_Lng
Andreas_Lng

Reputation: 1199

Evaluation using ROC curve

I have an image processing algorithm that takes three parameters and produces a binary image. To each image i have the ground truth. So i am able to calculate true/false positives/negatives. The resulting binary image depends on the 3 parameters. Each parameter can take values in an interval, let's say from 0.0 to 1.0. To find the best result i run my algorithm with each combination (1000 in total). And for each result i calculate the statistics.

Now i want to know which parameters lead to the best result. I know the ROC curve from algorithm engineering. Is this the way to go or is it better to compute the F1-score and take the parameters that leads to the highest score?

The roc curve would then look like this:

enter image description here

Upvotes: 0

Views: 79

Answers (1)

Calimo
Calimo

Reputation: 7959

You are misunderstanding what a ROC curve is.

One builds a ROC curve from one predictor (with one fixed set of parameters) by varying the decision threshold. This is the exact opposite of what you've done here, where you varied the parameters and likely kept he threshold fixed (0.5 maybe?).

What you could do is build 1000 ROC curves and select the best. Whether it would be appropriate in your case is an other question.

Upvotes: 2

Related Questions