Manually annotated characteristic descriptors: Measurability and variability

Abstract

In this paper we study the measurability and variability of manually annotated characteristic descriptors on a forensic relevant face dataset. Characteristic descriptors are facial features (landmarks, shapes, etc.) that can be used during forensic case work. With respect to measurability, we observe that a significant proportion cannot be determined in images representative of forensic case work. Landmarks, closed and open shapes, and other forensic facial features show mostly that the variability depends on the image quality. Up to 50% of all considered evidential values are either positively or negatively influenced by annotator variability. However, when considering images with the lowest quality, we found that more than 70% of the evidential value intervals in principle could yield the wrong conclusion.