I want to do facial recognition on wide varieties of images captured at various ages of my family members. Below are some of the questions I have.

  1. If a person uses glasses of different types, do I need to feed images with each glasses type in the training dataset ? What if the eyes are not visible after wearing sunglasses. ?
  2. There are many pics which ranges from childhood to current age of approx 30. Do I need to train my model with pics of various ages to get good accuracy ?
  3. Do I need to align the images of train dataset for better accuracy ?
  4. Do I need to crop the background of the images in the training dataset just to keep the faces OR background doesn't matter ?

The idea behind my requirement is to scan through all the images on my hard disk (around 100 GB images) and classify the images in different folders for each person. I know similar thing is done on iphone and other mobile phones but I wanted to do it for the offline images stored on my external hard disk.

I was exploring facenet model from google, opencv face recognition. But stuck in the first step of dataset creation.

  1. Glasses matter. So you will need those photos in the training dataset.
  2. Yes. You will need to feed images of all ages for multiple people. This will help the model generalize how people age
  3. Alignment should not matter
  4. Background should not matter
  • $\begingroup$ Thanks Suneet.. I am sure there must be some solution to handle glasses as people can wear any type of glasses and if we end up feeding each type in the training dataset then its a pain.. iphone handles this case.. not sure how to handle in offline files.. $\endgroup$ – Akash Jun 15 '20 at 4:56
  • $\begingroup$ Any reference blog/code which can help me implement this scenario ? $\endgroup$ – Akash Jul 11 '20 at 3:18

I would like add few points mentioned by Suneet Bhatia & you questions regarding iphone

  1. You have to add images with glasses to be recognize by it & definitely it will hit your accuracy. The way iphone face recognition works is different then facenet model. Iphone use face verification which is comparing one to one face & iphone not only use normal image with it use 3d structure of the face using IR camera.
  2. Added images of all age group image if your target audience of the same group otherwise it is not required.
  3. It does not matter if you don't align the face images.
  4. Always try to remove extra element from the image so try pass only face images into the network for better perform.

Problems like facial recognition, image search & semantic similarity are not actually a classification problem they are more like differentiate between 2 elements

  • $\begingroup$ Thanks @Swapnil. 1. So does that mean I need to include images of different glasses type ? 2. Basically I want to give input as few images of my son and I want to find all his images from the hard disk which has around 100 GB data. If I end up finding his images manually from the hard disk it will defeat the purpose of writing this code. No ? 4. Is there any automated way to do it or again I have to open up each image manually and crop it ? $\endgroup$ – Akash Jun 16 '20 at 5:22

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.