Like subset-based methods, the very first finite element versions of digital image correlation were closely related to the regular structure of images, as they were based on regular quadrilateral elements corresponding to an integer number of pixels. The use of unstructured meshes, to exploit the full potential of FE-DIC in structural mechanics, is now widespread. Most of the time, the formulation, the quadrature and the definition of the region of interest still rely on the pixels grid. In this paper, a formulation in the physical coordinate system and not in the image frame is proposed for 2D digital image correlation. In addition to a more precise definition of the region of interest, it allows the use of a more accurate quadrature rule. It is also shown that lens distortions can be successfully taken into account directly with such a formalism.