DIFFUSE

In many machine learning (ML) based models that can be applied to increase traffic safety, comfort and security, e.g. drowsiness, gaze and head pose estimation as well as authentication, the datasets necessary for training and verifying performance need to be huge. These machine learning models must be developed using a dataset with a large number of individuals with different ethnical characters in order to be robust, where the individual personal features must be diverse. The data can only create an approximation of reality and is often not globally applicable, e.g. data collected in Sweden may not be applicable in Asia, and this shortcoming extends to other aspects in the data. 

One of the methodologies for alleviating these issues is collecting new data, which is both costly and liable to the same problems as before. A second method is to generate data specifically for validation. However, a problem when generating images for validation purposes it that despite if one finds images where model performance is poor, it is not necessarily easy to tell what in the image led to this deterioration in performance.  

Our proposed solution to improve upon data generation is to use machine learning models to generate synthetical faces with diverse and controllable facial attributes. Furthermore, the developed models will be used to create an increase in explainability and control when validating authentication, head pose and gaze estimation models.