Face-Unlock

This is Face-Unlock repository of IvLabs and contains the implementation of Triplet Network and FaceNet with ResNet as the backbone architecture implemented from scratch to perform one-shot and zero-shot learning on different datasets.

Our work is categorized as following:

• [x] Triplet Network

  • [X] Triplet Loss on MNIST
  • [x] CNN on AT&T Dataset
  • [x] ResNet on AT&T Dataset

• [x] FaceNet

  • [x] AT&T Dataset
  • [x] LFW Dataset
  • [ ] Glint360k Dataset

• [ ] Real-Time Face Recognition

  • [ ] Hosting Web based implementation
  • [ ] Integrating with Rasberry Pi

Datasets

• The AT&T face dataset

  There are 10 different images of each of 40 distinct subjects.

  Dataset Statistics

  1. Color: Grey-scale
  2. Sample Size: 92x112
  3. #Samples: 400

• Labeled Faces in the Wild

  Labeled Faces in the Wild is a public benchmark for face verification, also known as pair matching.

  Dataset Statistics

  1. 13233 images
  2. 5749 people
  3. 1680 people with two or more images

ResNet

A part of this project was also to understand and implement Residual Networks from scratch which can be found in model.py

References

• Deep Residual Learning for Image Recognition
• Understanding and visualizing ResNets
• FaceNet: A Unified Embedding for Face Recognition and Clustering
• Triplet Loss and Online Triplet Mining in TensorFlow
• facenet-pytorch-glint360k