As the promise of Virtual and Augmented Reality (VR and AR) becomes more realistic, an interesting aspect of our enhanced living environment includes the availability - indeed the potential ubiquity - of scannable markers. Such markers could represent an initial step into the AR and VR worlds. In this paper, we address the important question of how to recognise the presence of visual markers in freeform digital photos. We use a particularly challenging marker format that is only minimally constrained in structure, called Artcodes. Artcodes are a type of topological marker system enabling people, by following very simple drawing rules, to design markers that are both aesthetically beautiful and machine readable. Artcodes can be used to decorate the surface of any objects, and yet can also contain a hidden digital meaning. Like some other more commonly used markers (such as Barcodes, QR codes), it is possible to use codes to link physical objects to digital data, augmenting everyday objects. Obviously, in order to trigger the behaviour of scanning and further decoding of such codes, it is first necessary for devices to be aware of the presence of Artcodes in the image. Although considerable literature exists related to the detection of rigidly formatted structures and geometrical feature descriptors such as Harris, SIFT, and SURF, these approaches are not sufficient for describing freeform topological structures, such as Artcode images. In this paper, we propose a new topological feature descriptor that can be used in the detection of freeform topological markers, including Artcodes. This feature descriptor is called a Shape of Orientation Histogram (SOH). We construct this SOH feature vector by quantifying the level of symmetry and smoothness of the orientation histogram, and then use a Random Forest machine learning approach to classify images that contain Artcodes using the new feature vector. This system represents a potential first step for an eventual mobile device application that would detect where in an image such an unconstrained code appears. We also explain how the system handles imbalanced datasets - important for rare, handcrafted codes such as Artcodes - and how it is evaluated. Our experimental evaluation shows good performance of the proposed classification model in the detection of Artcodes: Obtaining an overall accuracy of approx. 0.83, F2 measure 0.83, MCC 0.68, AUC-ROC 0.93, and AUC-PR 0.91.