This paper addresses the problem of foreground extraction using active illumination and graph-cut optimization. Our approach starts by detecting image regions that are likely to belong to foreground objects. These regions are constituted by pixels where the difference in luminance for two differently illuminated images is large. The foreground objects are segmented by graph-cut optimization using those regions as a seed and using a energy function based on probability distributions derived from both input images and their difference. Several light sources and different illumination schemes can be used to mark the foreground. Our method has only two scalar parameters which can be set once for a wide variety of scenes.