Should I use the default model in the deepface package or fine-tune it to fit with my data for face recognition?

42 Views Asked by At
  1. I have 1000 classes, each containing 10 images. For face recognition with my dataset, would it be more beneficial to use a pretrained model or to perform fine-tuning?

  2. Should I retrain from scratch, and if so, what techniques should I use for data augmentation?

1

There are 1 best solutions below

0
sefiks On

It depends! For instance, the most dominant 2 models in deepface's portfolio, VGG-Face was trained with 2.6M and FaceNet was trained with 200M training images. If you are going to train these models with more training data, the answer is yes, your fine-tuned model would most probably overperform than pre-trained models. Notice that these models were trained with very powerful GPUs and it taken days! So, using pre-trained models would be wiser for most cases.