CANUPO: What are good values for Balanced Accuracy and Fisher Discriminant Ratio for training classifier?
Posted: Sat Mar 11, 2017 7:37 pm
Hello!
I am in the process of training classifiers for vegetation and ground point separation. I have trained a classifier that samples 1000 points from each class with steps set at 1 to 10. The result of this was 0 points falsely classified for each class, a Balanced Accuracy of 1, and a Fisher Discriminant Ratio of 28.67. I read Brodu and Lague's paper on 3D Terrestrial Lidar Data Classification for Complex Natural Scenes. I understand that the Balanced Accuracy is related to the performance of the classifier and that the Fisher Ratio is to assess the class seperability. However, I don't understand what values I should be trying to achieve for these? I'm assuming that the Balanced Accuracy will always be between 0 and 1 and the closer to 1 the better (similar to Anova?) But a value of 28.67 for the Fisher Ratio is confusing to me.
Any help is greatly appreciated!
I am in the process of training classifiers for vegetation and ground point separation. I have trained a classifier that samples 1000 points from each class with steps set at 1 to 10. The result of this was 0 points falsely classified for each class, a Balanced Accuracy of 1, and a Fisher Discriminant Ratio of 28.67. I read Brodu and Lague's paper on 3D Terrestrial Lidar Data Classification for Complex Natural Scenes. I understand that the Balanced Accuracy is related to the performance of the classifier and that the Fisher Ratio is to assess the class seperability. However, I don't understand what values I should be trying to achieve for these? I'm assuming that the Balanced Accuracy will always be between 0 and 1 and the closer to 1 the better (similar to Anova?) But a value of 28.67 for the Fisher Ratio is confusing to me.
Any help is greatly appreciated!