Questions on locating and characterizing facial features in C#

[btb4198]

TL;DR Summary

I have some questions on locating and characterizing facial features in C#. facial features like Eyes, Noses, Mouths, ears, pupils, eyebrows, and hair.

I have some questions on locating and characterizing ( labeling) facial features in C#. Facial features like: Eyes, Nose, Mouths, Ears, Pupils, Eyebrows, and Hair.
My objective in this project, is to learn more about neural networks. I am very new to neural networks.
I have a lot of Images like this:

They are all 5184 X 3456 but of different people, at different Focal lengths and with different backgrounds.

for example:






I want to design and code a Gans network in C# to locale and label the pixels with each facial feature.
I have been doing some research and I learned about Filter banks and how they are an an important tool for object classification in images.
So I started looking into filter banks and I learned about the VGG-16 model. The VGG-16 model is a deep convolutional network that was proposed by Simonyan and Zisserman in 2014. It is made up of 16 layers, 13 of which are convolutional. The VGG-16 model has been pre-trained on a large dataset and is widely used for image classification tasks.

Question 1, is this the best way to go for my application?

I learned there are two C# libraries that already implemented the VGG-16 model.
ConvNetCS and accord-framework

Question 2 , Has anyone used either of these libraries before?

Question 3, which one is the best for my application and why?

Last question, during my studying of deep learning, I learned there are a few factors to consider when choosing the correct filters for a convolutional network:

- The size of the input data
- The type of input data
- The desired output

The size of the input data is important because it determines the size of the filters.
For example:
If the input data is 224 x 224 the convolutional filters should be 3 x 3 or 5 x 5 , but I have not been able to find an equation to relate the input data size to the needed filter size.

Question 4, is there an equation to relate the input data size to the needed filter size?

Also if I have said anything wrong in this post, please correct me, I am completely new to this subject.

 

[Jarvis323]

Never done anything like this, but the first one only has pretrained models for general object recognition and scene recognition. The second one says it has a model for detecting and analyzing facial features, so it seems a better fit.