TFCV is my repo for experimenting with Computer Vision models in Tensorflow 2. The goal is to have everything simple and searchable.

• TO-DO:
  • create WiderFace dataset / TF Records using tfds
  • pick apart and implement RetinaNet repo for understanding
    • load backbone model from tf.keras.applications for transfer learning
    • create FPN
    • create context layers for classification and regression for final output
    • work on anchor boxes
    • image augmentations and incorporate w/ anchors for preprocessing (need to make sure crop has at least one box)
    • figure out loss
    • understand inference differences (non-max supression, etc.)
    • modularize for tensorflow 2.2 (can now change train_step and predict_step) (currently working on)
    • add callbacks
  • create notebooks for each piece w/ visualizations and understanding
  • run model w/ hyperparameters from RetinaFace modified for tf==2.2 and if good enough convert to tflite
  • take trained model and fine tune with ArcFace loss for face embeddding
  • convert that to tflite and get working on RaspberryPi
  • add new backbone or object detection models
  • add distributed computing