I made a face recognition system that takes 300 images of individuals. As of now there are probably 5 persons in the local dataset that I have which contains 1,500 images. The prediction values if it is on multi faces play around 50-90% from a live feed. I would like to compute the accuracy of the face recognition in multi faces. I saw threads about TP? or True positives. And I am having difficulty understanding it. Is there like a layman’s term for computation of the accuracy? I am sorry if I sound really dumb, which I am. For added reference if needed I did not download any other dataset.
here, it’s as simple as:
accuracy = correct_predictions / all_predictions
also, setting up a confusion matrix, where you have “predicted” id’s on the columns and “expected” ids on the rows, will be quite useful:
conf = np.zeros((nclasses, nclasses))
#for each prediction:
conf[predicted, expected] += 1
...
print(conf)
which should result in something like:
20 0 0 3 0
2 10 0 0 0
0 0 16 0 0
1 2 0 20 0
0 0 0 0 20
(“true positives” are on the diagonal)
those terms are hard to apply in the multi-person, identification setup here, e.g. there are no “false negatives” at all, while it makes a lot of sense for a verification task (image a & b show the same person ?)
unclear, what do you mean, please ?