Issues with reproducing some accuracies illustrated in insightface/model_zoo github

Hello,

We have issues with reproducing accuracies illustrated in insightface/model_zoo github through the following table :

We downloaded the associated .onnx files from the link in the last column of the table. We used a python script and loaded the .onnx files using the "load" method of onnx package. We also used "get_model" method of "insightface.model_zoo" package. In both cases, we found different results from the table. In both cases, we obtained different results from the table for LFW dataset (80.4% vs 99.55% for VGG2).

Do we need to preprocess the .onnx file in order to get the accuracy you show in the table?

Best regards,

Start by reproducing the reported LFW result with the downloaded .onnx files, comparing the Python script's onnx loading path with insightface.model_zoo.get_model. Check the table's VGG2 entry and the stated preprocessing assumptions; done means the accuracy discrepancy and required preprocessing are documented or the affected model information is corrected.

Written by the indexing model from the issue text.