Accurate and fast building feature extraction is of great significance for construction of smart cities. Polygonal features and linear features are the most fundamental geometric information, which can accurately reveal building structural characteristics. This study first conducted comparative analysis of surface feature detection using three distinct approaches: Euclidean clustering, random sampling consistency algorithm and regional growth method, followed by quantitative performance comparison. For edge characterization, the linear features were detected by normal vector and approximate curvature estimation,respectively, and the effects of them were compared and analyzed. Experiments show that the regional growth method demonstrated superior precision of surface feature extraction compared to other methods, whereas the approximate curvature estimation enhanced continuity of smoothness of linear characteristics. The methods adopted in this study for accurate and efficient detection of building features has considerable application value for subsequent reconstruction of building models.