Experiments with face image landmarking and transfer learning using Keras inspired by the blog posts:
- https://towardsdatascience.com/face-landmark-detection-with-cnns-tensorflow-cf4d191d2f0
- https://machinelearningmastery.com/how-to-use-transfer-learning-when-developing-convolutional-neural-network-models/
Download face images with marked landmarks dataset at https://www.kaggle.com/drgilermo/face-images-with-marked-landmark-points/ and store them in "datasets" in the root directory.
Ensure that consistent versions of the various packages required for the code are used by generating a virtual environment and installing pip packages (python3) listed in requirements.txt:
pip3 install -r requirements.txt
Train, test and visualize
python3 train_test_detector.py
- Make it work... accuracy ~ 1e-4.
Create loss, accuracy summaries.- Retrain entire model, the VGG16 weights may not be good for grayscale input.
- Try other base models.
Store trained model for later use.