A Robust Performance Evaluation Metric for Extracted Building Boundaries From Remote Sensing Data

Abstract

Various methods for automatic building extraction from remote sensing data including light detection and ranging (LiDAR) data have been proposed over the last two decades but a standard metric for evaluation of the extracted building boundary has not been found yet. An extracted building boundary from LiDAR data usually has a zigzag pattern with missing detail, which makes it hard to compare the boundary with its reference. The existing metrics do not consider the significant point (e.g., corner) correspondences, therefore, cannot identify individual extralap and underlap areas in the extracted boundary. This article proposes an evaluation metric for the extracted boundary based on a newly proposed robust corner correspondence algorithm that finds one-to-one true corner correspondences between the reference and extracted boundaries. Assuming a building has a rectilinear shape, corners and lines are first detected for the extracted boundary. Then, corner correspondences are obtained between the extracted and reference boundaries. Each corner has two corresponding lines on its two sides that ideally are perpendicular to each other. The corner correspondences are finally ranked based on their distance, angle, and parallelism of corresponding lines. The metric is defined as the average minimum distance d_{\mathrm{avg}} from the extracted boundary points to their corresponding reference lines. Extralap and underlap areas are identified by comparing the point distances with d_{\mathrm{avg}}. In experiments, the proposed metric performs more realistic than the existing metrics and finds the individual extralap and underlap areas effectively.