1

we are using caffe to do the classification of images, there two classes. We've done the training and testing and get the accuracy. Now I want to find the false positive images from the test dataset. Is there anyway I can do this? Thanks!

ccatman
  • 11
  • 2

2 Answers2

0

The easiest way I know of is to run them through classification -- as a separate run, rather than as part of the training process -- and see which ones didn't come out as expected.

Prune
  • 76,765
  • 14
  • 60
  • 81
  • Thank you for answering the question. But I cannot fully understand what is the meaning of separate run. I am trying to find out the false positive images in the test set. Could you elaborate more about it? – ccatman Mar 21 '17 at 16:29
  • Look up how to actually *use* your trained model: how to classify new inputs. This is also called "prediction" or "scoring". Instead of using new inputs, use the testing set. Compare the returned predictions to the ground truth (original classifications). – Prune Mar 21 '17 at 16:32
  • Thanks, ill go try – ccatman Mar 23 '17 at 15:44
0

You need to label your test data somehow, to know if your positive prediction matches the positive test sample, if not then it should be considered a false positive, also when you have these metrics for all your test data, you can calculate the f1score and some other performance variables:

accuracy = (true_negative + true_positive)/total_samples
recall = true_positive / (true_positive + false_negative)
precision = true_positive / (true_positive + false_positive)
f1score = 2*((precision*recall)/(precision+recall))
Juan Zapata
  • 1,818
  • 14
  • 19