Image outlier detection (OD) is an essential tool to ensure the quality and accuracy of image datasets used in computer vision tasks. Most existing approaches, however, require a set of in-distribution data for training prior to outlier prediction. The quality and quantity of the data can influence the resulting performance. Thus, selecting a suitable in-distribution set often requires considerable effort. In this work, we propose RANSAC-NN, an unsupervised image OD algorithm designed to detect outliers within contaminated sets in a one-class classification fashion. Without any training, RANSAC-NN performs favorably in comparison to other well-established methods in a variety of OD benchmarks. Furthermore, we show that our method can enhance the robustness of existing OD methods by simply applying RANSAC-NN during pre-processing.