AGATHA: Face Benchmarking Dataset for Exploring Criminal Surveillance Methods on Open Source Data

This paper outlines the development of a group of benchmarking datasets, using open source data, that provide useful information to improve the use of computer methods in tasks such as criminal control and surveillance. We aim to build a golden standard reference for developing computer vision systems capable of recognising possible criminal actions and, in consequence, given to the criminal investigation police the capacity to prevent them through faster action. We have explored, developed and fine-tuned Deep Learning models that demonstrated a statistical high performance in pattern recognition tasks such as: face detection (high accuracy specifically in gender detection and age interval determination), face recognition (people tracking and recognition on video sequences) and object recognition (i.e. cars and guns). In this stage, we present the first iteration through this set of datasets, where we present a dataset for age and gender recognition.