[3] Dlib Converter 2.0: Better Conversion Based on Classification Library Data

The face has a great deal of variability. When starting with a 2D source input, it’s sometimes tricky to tell, even with the human eye, the position a head is looking. To complicate things, we don’t always know the exact angle of the camera. These factors all play a part in pose estimation, normalization and head rotation.

Head rotation was initially thought of as a simple problem to solve. We use solvePnP and for the most part, it gives us some results. But the problem is that the results are based on an estimation from a single frame. The next frame is a unique and special estimate that might not be compatible with the previous frame.

WIP - To Be Continued…