Complete Convolutional Neural Networks Environment for Computer Vision Problems With Nvidia Drive AGX Xavier

Development of Convolutional Neural Networks (CNN) models with supervised learning in computer vision tasks can be a demanding process. Multiple components lead to creation of the final model including data acquisition, image labeling, data preprocessing, model training, testing and key performance indicators computation. In general these tasks require significant hardware resources from both storage and computational power points of view. In this paper we propose an alternative lighter environment for the entire process based on the Nvidia Drive AGX Xavier development platform. Using methods for automatic image labeling and fully automated environments of CNN models training and testing, this method is much easier to setup and use. Additional advantages of portability to use it in real-life scenarios makes it a very flexible development environment. We present experiments performed for driver monitoring computer vision tasks with similar applicability in other detection problems.