Description
Hi,
Thanks for your great work and sharing of the code on these two papers !
It takes me days to read the paper and go through the repository and I have a few questions:
(2)
Do you have the code for training the features (asian_females, asian_males, black_females, black_males, indian_females, indian_males,...). Since I have a hard time finding something like train.py (e.g. the loss function and training process).
(I suppose the released code is mainly on image pre-processing and result analysis)
(Since BFW dataset is not as large as other face dataset and it may possible for me to train it from scratch on one GPU)
(3)
I am little confused about how the BFW is used in two papers, as I understand:
in paper Face Recognition: Too Bias, or Not Too Bias? , the train and test model are as follows:
train: CASIA_webface trained using Sphereface loss
test: LFW
where does BFW dataset not used in training in this set of experiments?
in paper Balancing Biases and Preserving Privacy on Balanced Faces in the Wild the train, test model are as follows:
tain:
(1) MS1M trained using Arcface loss --> to get 512-dim embedding (f_in in Fig.6)
(2) BFW dataset is used to train the encoder and two classifiers in Fig 6
test: 4-folds used for training and 1-fold used for testing (using the best threshold chosen)
is that right?
(4)
There are some difference from "bfw-v0.1.5-datatable.csv" and the TABLE-2 in paper 2:
for example: there are 921379 records in TABLE-2 while ther are 923898 records from the csv file?
and there is no "{dir_meta}thresholds.pkl" file.
Thanks for your time and any help would be appreciated !