This README shows how to set up the downloaded face images from the LFW dataset for face verification evaluation.
Download the deep-funneled (roughly-aligned) images from LFW: http://vis-www.cs.umass.edu/lfw/lfw-deepfunneled.tgz
. Extract these at some local folder, referred to henceforth as LFW_DIR
.
NOTE: all scripts are to be executed from the project root directory, not the sub-folder where this README file is located.
The LFW database provides DevTrain and DevTest splits as validation sets for developing code without overfitting to the 10 folds used in the final evaluation. The DevTest pairs are saved at ./lfw/data/pairsDevTest.txt
Run python ./lfw/eval_lfw.py -d LFW_DIR -m MODEL_PATH --fold 0
to get AUC score for verification on the dev test split. The number should be around 0.9989. Setting the --fold
flag to 10 will evaluate on the full 10 folds of LFW test set. Note, ROC curves and AUC metrics are reported (not EER).