Main subject detection (MSD) refers to the task of determining which spatial regions in an image correspond to the most visually relevant or scene-defining object(s) for general viewing purposes. This task, while trivial for a human, remains extremely challenging for a computer. Here, we present an algorithm for MSD which operates by adaptively refining low-level features. The algorithm computes, in a block-based fashion, five feature maps corresponding to lightness distance, color distance, contrast, local sharpness, and edge strength. These feature maps are adaptively combined and gradually refined via three stages. The final combination of the refined feature maps produces an estimate of the main subject's location. We tested the proposed algorithm on two extensive image databases. Our results show that relatively simple, low-level features, when used in an adaptive and iterative fashion, can be very effective at MSD.