Genuine and Imposter Scores

I'm following the matlab example for "Transfer Learning Using AlexNet" and I'm trying to figure out how to calculate the genuine and imposter score distribution. If anyone has any help on how to do it, that would be great. Not looking for the exact code, just trying to get a better idea on how I can do it myself. Kind of struggling with it. Thanks.

Answer 1

[Prasanna]

Hi Christian,

To calculate genuine and imposter score distributions, you typically need to extract feature vectors for each input sample (such as images) from a layer in the AlexNet model—usually a fully connected layer like 'fc7' or even the output of the last ReLU activation before classification.

Once these feature vectors are extracted, you compute a similarity score (such as cosine similarity or Euclidean distance) between pairs of features. Genuine scores are the similarities between feature pairs from the same class or identity, whereas imposter scores are from different classes or identities. To do the same, you can:

• Use ‘activations’ in MATLAB to extract features from the desired layer.
• Group the extracted features by class labels.
• For each class, compute similarity scores between all intra-class (same identity) pairs — these are the genuine scores.
• Then, compute similarity scores between all inter-class (different identity) pairs — these form the imposter scores.

For more information regarding some of the functions mentioned, you can refer to the following documentations:

• activations: https://www.mathworks.com/help/deeplearning/ref/seriesnetwork.activations.html
• cosine similarity: https://www.mathworks.com/help/textanalytics/ref/cosinesimilarity.html

Hope this helps!