Gender accuracy in lfw is very low
xudou3 opened this issue · comments
Hi, i test this model with a little bit modify image_emotion_gender_demo.py to classify gender laber on data set lfw(http://vis-www.cs.umass.edu/lfw/) and the accuracy for female is 82% and for male is just 69%.
What I have done to the code is just make image_path in a for loop to read all file in a floder. And there was an error occured when load rgb image in utils/inference.py in function load_image with unmatched parameters so I fix it.
I am using python2.7 and tensorflow1.8.0 with GPU support.
Hi thank you for your comment. I am actually impressed on how high
it is. I have discussed previously (and within the paper) that the "sex" classifier is biased given the data in which was trained on.