Thanks for your interest!!
1. One of the important concepts to understand in this project is that the predicted landmarks (output of the neural network) have values between 0 and 1. So, multiply them by the dimension of the input image and take the integer value of the result. This will give you the true landmarks in terms of pixel locations in the cropped faces. Now, add the offset of the detected face in the original image to the landmarks to get the global location for each set of landmarks.
2. Repeat this for every face detected in the original unseen image. If you do 1 right, you will get 2 and 3 easily. I have included the code for it as well. Run it locally. If it's not broken, it should work fine!