A heuristic model based on Laplacian Standard deviations. Learning based models are yet to come.
For StdLaplacian, 3.8 seems to be a good threshold.
Faces with StdLaplacian less than 3.8 can be considered blurry.
The script blur_test.py is an example on how to do a grid search for the ideal threshold on an image folder.
- build a blur face datasets
- train a lightweight model for blur detection