facenet
facenet copied to clipboard
Poor Performance for dark complexion faces
I Am Trying to build a face verification system, and my target dataset is of african people images. But not getting any promising results using the base model. Two potential reasons for that could be:
-
Embeddings getting generated from id card images are not good(because of poor image quality maybe)
-
Since the face images have a bit more black complexion, and hence the model is not able to create good embedding on those face images.
Can someone share their thoughts as to how i can tackle the problem, and if there is any african face images(or any other images dataset with black complexion images available). I have ample amount of dataset available, but have only two images for each class(one from id card and the other as a normal selfie image) and hence not sure if those can be used for fine-tuning the model.
I believe the second problem can be taken care of by fin-tuning the base model on appropriate dataset, but not sure how to tackle the first issue.
Appreciate any help and suggestions.