Computer Vision applied to Historical Air Photos: The Registration and Object Detection Challenge

Sebastian Zambanini, Fabian Hollaus, Robert Sablatnig


This paper addresses the problem of automatically analyzing aerial photos taken during World War II air strikes. The goal of this work is to locate unexploded ordnances (UXOs) for risk assessments, enabled by the registration of the historical air photos to modern-day satellite images and the detection of military objects (e.g. bomb craters or trenches). The work is part of the DeVisOR project which aims at supporting the tedious task of creating UXO surveys in a semi-automatic manner by means of powerful image analysis methods and interactive visualization techniques.

In this paper we focus on the image analysis part and present the specific challenges that arise when working with this kind of data. For registration, the strong image changes caused by time spans of around 70 years hinder the reliable identification of correspondences between the old and new images, especially in non-urban areas. In combination with the generally low image quality of the old aerial photos and the appearance variations caused by illumination changes, a straightforward solution based on standard algorithms using keypoint matching and sample-based transformation estimation does not exist. The same problem appears for the detection task, which is additionally impeded by the absence of large amounts of training data. Consequently, innovative solutions are required that are tailored to the specific conditions of the problem.