search

sajjjadayobi / FaceLib

Face Analysis: Detection, Age Gender Estimation & Recognition
MIT License
300 stars 52 forks source link

some questions #26

Closed JH-Lam closed 2 years ago

JH-Lam commented 2 years ago

Hi , thanks your great job on this open project. there are some questions I confused , please give some clues on them:

  1. which sub dataset is used for training, the original wild or cropped? anyone is viable?
  2. why dont you use MoibleNet but ShuffleNet in Age and gender model? appreciate your effort again!
sajjjadayobi commented 2 years ago

Hi thank you, sorry I'm late. Answer to your first question: The original WILD Second one: At that time I didn't know MoibleNet is a better architecture, but If I wanted to train the model again I would use Efficientnet-b0

JH-Lam commented 2 years ago

thanks. TBD: Do you think it's more reasonable to use cropped dataset to train the model so it'll gain higher accuracy?

sajjjadayobi commented 2 years ago

Ye I think that's a reasonable thought 🤔

JH-Lam commented 2 years ago

sorry I'm a bit later. Yes I just wonder what the landmarks are if an image is not a face when labels the training set? for instance, image1 is a cat, so it's class 0 ( class 1 belongs to face), and its landmarks should be all 0 ,-1 or others? what differences b/w these as to performance? thanks a lot!

sajjjadayobi commented 2 years ago

Great question, but I never thought about or tested this one. I don't know what will happen. have a try at it p.s. give a star if the project was helpful

JH-Lam commented 2 years ago

You mean your training set are all about facial scenarios but no negative examples( eg. only a car on a road, flowers in the bottle etc) ? it seems weird since the model will still predict a face w/ landmarks even if it predicts on images without faces

This problem will not raise if splits one model to two: one for classifier , another is for regression on the positive result of former. (ps : starred it)