How to replicate the result of ID similarity distribution on your ageDB-400

[samfxbhgh]

I would like to replicate the ID similarity distribution results from your ageDB-400 dataset, but I noticed a potential discrepancy with the ID similarity distribution shown in Fig. 4(b) of your paper. Specifically, the mode of my test results is less than 0.8, whereas in your paper, it appears to be above 0.8.

Here is the procedure I followed:

I used the test data and the InsightFace API as mentioned in the repository to extract the ID embeddings (det_size=(480, 480)). Then, for each embedding, I generated 5 results and used the same method to obtain the ID embeddings of these generated results. I calculated the cosine similarity between the original and the generated embeddings. Finally, I used sns.kdeplot to visualize the ID similarity distribution based on the resulting 500x5 data.