We developed a method to evaluate the accuracy of segmentation algorithms. Oversegmentation, undersegmentation, missing and spurious labels may all appear concurrently in machine segmented images. Segmentation algorithms make systematic errors and have different optimal operating ranges. Existing methods of segmentation evaluation do not evaluate these details. Our method, based on multiple feature recovery, reports systematic errors and indicates optimal operating ranges of features, besides measuring overall errors. A knowledge of the magnitude and type of errors can be used for tuning or selecting segmentation algorithms. Although our method was developed for CT scanning for security, it is applicable to other fields, including medical imaging, where multi-object feature recovery, non-uniform costs and a knowledge of optimal operating ranges are helpful.