A revisited visual-based geolocalization framework for forensic investigation support tools

Visual-based geolocalization (VBL) is a well-studied computer vision topic, and has undergone a dramatic improvement in the last decade by the emergence of deep-learning techniques. However, previous studies about VBL have not mainly focused on forensic purposes. In this study, we revisit the traditional VBL framework by replacing its elements with state-of-the-art machine learning techniques, and discuss what points of a VBL framework are important for investigation support tools. Based on this revisit, we claim that there are two essential points for investigation support tools. One is maintainability that results from the ease of replacing the implemented algorithms with more appropriate algorithms. The other is the interpretability of bottlenecks in the current system. These two advantages are specific and essential to forensic science where practicality is important.