In recent years, the LIDAR technique has undergone fast development. The increasing access and operating ability caused a growing interest in 3D processing of data acquired by LIDAR. One of the main tasks of geo-information modeling is to create virtual city models. As the available commercial softwares require a high level of user interactivity, the crucial issue of modeling is its automation. There are four main steps that comprise virtual building extraction. One of them, building point cloud segmentation, appears to be the core part of the whole modeling process. Segmentation allows partitioning of a data set, that contains points biased by random and gross errors, into smaller sets which represent different planes. This arises from the fact, that buildings are formed by a combination of planes in 3D space. The paper presents an analysis of two algorithms that are most commonly applied to segmentation: RANSAC and plane growing. The latter is modified, taking into consideration topology between points. The essential information about both algorithms is presented. Numerical tests based on synthetic and real laser scanning data are executed. It is inferred from the experiments that the RANSAC algorithm features short time performance for simple models. However, at times it merges different objects lying in the same plane. The algorithm is suited well for segmentation of standard roofs that contain small number of elements. The plane growing algorithm is more suitable for more complicated models. It separates different objects situated in the same plane. Time performance depends mostly on the number of points within a data set; it is not affected by the number of identified planes.