The post describes the steps that were made in order to speed-up Face Detection.

Rewriting of MB-LBP function into Cython

The MB-LBP function is called many times during the Face Detection. For example, in a region of an image that contains face of size (42, 35) the function was called 3491 times. The sliding window approach was used. These numbers will be much greater if we use bigger image. This is why the function was rewritten in Cython. In order to make it fast, all the Python calls were eliminated and the function now uses nogil mode.

Implementing the Cascade function and rewriting it in Cython

In the approach that we use for Face Detection the cascade of classifiers is used in order to detect the face. Only faces pass all stages and are detected. All non-faces are rejected on some stage of cascade. The cascade function is also called a lot of times. This is why the class that has all the data was written in Cython. As opposed to native Python classes, cdef classes are implemented using struct C structure. Python classes use dict for properties and method search which is slow.

Other additional entities that are needed for cascade to work were implemented using pure struct C structure.

New module

For the current project I decided to put all my work in skimage.future.objdetect. I did this because the functions can be changed a lot in the future. The name objdetect was used because the approach that I use will make it possible to detect not only faces but other objects on which the classifier can be trained.