search

shamangary / SSR-Net

[IJCAI18] SSR-Net: A Compact Soft Stagewise Regression Network for Age Estimation
Apache License 2.0
584 stars 146 forks source link

Accuracy question about gender estimation #7

Closed mayidu closed 6 years ago

mayidu commented 6 years ago

hello, I use SSR-Net to estimate gender with wiki pre-model, but the value obtained does not seem to be a confidence， the values of males are greater than 1, and the values of females is not less than 0.5, mostly is 0.9999.... , so there are many errors with 0.5 as threshold. Is there a problem with my test? Thank you!

shamangary commented 6 years ago

Hello. In the real-world experiment, we found that SSR-Net for gender recognition may only detect long hair for determining the gender. Therefore a girl with short hair may also be recognized as a male.

This is caused by the dataset and the cropped range in face detection level. For example, we choose 40% wider region in the face detection region. It is suggested by the previous "age estimation" method. However, you could rebuild the dataset by setting the extra range smaller (or even smaller than the original face detected region) to avoid the hair interference. Or you may simply change a dataset which contains more short hair females.

I will retrain the gender model by such setting in the future. Thank you!

mayidu commented 6 years ago

I understand, thank you very much and look forward to your retraining gender model.