In this project, we want to feed in images of shape 224*224*3 and be able to predict the joint coordinates for the body in the image. That is, we want to be able to predict 14 numerical values (7 (x,y) pairs). Therefore, we modify the final Dense layer to have 14 neurons and use ReLU activation instead of sigmoid.
Training a deep learning model such as VGG16 can take up to a week on a local machine. This is a lot of time. An alternative to this in our case is to use the weights of a trained VGG16 model through transfer learning.
We will do this with the help of the applications module in Keras that we imported in the beginning, along with the other imports.
In the following code, we will load part of the VGG16 model ...
