Original article can be found here (source): Deep Learning on Medium
Real-time Facial Surface Geometry from Monocular Video on Mobile GPUs
Just from one image, → we are able to create a mesh map → that are pretty dense → a lot of key points in the mesh.
Computervision → landmark detections were done again and again → these are tracked since we are now able to solve a lot of different problems.
Creating realistic mesh is hard to do → and quite the challenge. (there were 468 points arranged → that are quite higher than I expected → but we are able to create smooth mesh map → from this method). They even have model versions of CPU → for lightest models → they are able to survive on CPU.
So the method goes → the image comes into the camera → and next, we are going to make simple face predictions. (then we are going to use those data to do more processing).
These small little tricks → which are not the main contributions → actually have a large effect on the performance of training. (and additional data augmentations were done → this is a good and reasonable thing → since this model is going to work on the real-world setting).
They used a fairly reasonable residual connected network → quite the standard choice. (these research methods → they used different knowledge from other studies).
MAD → was used → mean absolute distance
So those glasses were AR projects.