Abstract. Introduced seven years ago, the correlogram is a simple statistical image descriptor that nevertheless performs strongly on image retrieval tasks. As a result it has found wide use as a component inside larger systems for content-based image and video retrieval. Yet few studies have examined potential variants of the correlogram or compared their performance to the original. This paper presents systematic experiments on the correlogram and several variants under different conditions, showing that the results may vary significantly depending on both the variant chosen and its mode of application. As expected, the experimental setup combining correlogram variants with boosting shows the best results of those tested. Under these prime conditions, a novel variant of the correlogram shows a higher average precision for many image categories than the form commonly used.