HyperFace

Contains the implementation of HyperFace: A deep multi task learning framework for facial recognition, landmark detection, pose and gender detection

TODO :

• Implement HyperFace-AlexNet

  • Selective search
  • Iterative region proposal (IRP)
  • Landmark based - Non Maximum Suppression (L-NMS)
  • skip training RCNN on imagenet and directly train on region proposals from AFLW dataset
  • Copy weights to the hyperface model and retrain it on AFLW dataset

• Implement HyperFace-ResNet101

  • Uses keras.application to load model weights and retrain R-CNN
  • Re-write the model architecture for resnet and carry out training

• Experimental stuff

  • Find better loss functions/ Implement custom loss functions
  • Find better optimizer
  • Decide on using ReduceLROnPlateau or not, look for useful callbacks

• Future Plans

  • Implement using ResNet
  • Divide the code into seperate files
  • support command line arguments

• Known issues

  • openCV's imread function reads in BGR instead of RGB, skimage ioread is comparatively slower than openCV, matplotlib shows wrong image plots because of BGR mode
  • Some images are bw by default so proper dimension conversion can not be done at all images (at the moment those images are skipped to save me from headache)
  • Processing all images uses up 14gb+ of RAM, so try and implement some other way because I plan on implementing data augementation in future to increase face detection accuracy
  • No official keras implementation of ResNet101, this code relies on experimental code(ro be changed in next version update)

Current model archietecture (AlexNet) is shown below -