An attempt to create and understand the working of snapchat filters using facial landmark tracking
- Also checkout my work on facial landmark tracking Facial-Keypoints
- Image --> CNN --> Output Co-ordinates of the landmarks
- Then use these co-ordinates to place filters on the image
- Since the data set only contain image of size 96,96,1
- Hence haar-cascade is used to first isolate the face and that is passed into model
- The model detects the landmarks and passes it to
filter.py
which will then apply the filter to the image - Then image processing is done to re-adjust the image in the vedio
vedio.py
runs the network on vedio
- Time everything and see what
- Use Numpy Operation Instead of Loops for applying filters on the image
- Write an argument parser for command line
- Add filters for mouth also! (using the lip tracking mechanism!)
- Train a face detection algorithm instead of haar-cascades
- If haar-cascades is taking significant time
- If calculation label points takes time , then save the points for a particular person and only detect if a new person is found
- Make a youtube vedio and blog for it!
- Implement Optical Flow stuff to give a 3D View?