Resumen
In this paper we propose hand-crafted physical features for video that can be used in a wide range of different regression and classification problems in video processing. For evaluation of resulted set of features we consider video genre classification problem among four classes: animation, drone video, computer game and sports. In this work we describe an automatic approach for video dataset creation, its augmentation and anomaly detection. In order to arrange the experiment we create dataset from 14271 samples, having a minimal number of samples per class is equal to 2700. Using gradient boosting we trained decision tree ensemble model with average precision and recall equal to 86.15% and 86.12% on test dataset. Also, other machine learning methods such as linear regression, naïve Gaussian classificator, support vector machine and random forest demonstrated worse results. The most relevant physical video features are two blur metrics using Laplacian operator and based on re-blur effect.