TL;DR
This paper introduces a novel method for automatically generating vectorized 3D building models from orthophotos and digital surface models (DSM). By deforming building contours to align with dominant directions, the approach creates regularized and compact models suitable for GIS applications, outperforming existing methods in accuracy and corner preservation.
What Problem Does This Paper Address?
Creating accurate and efficient 3D models of urban environments is crucial for applications like urban planning, autonomous driving, and virtual reality. While remote sensing technologies provide abundant 2D and 3D data, automatically extracting semantic, geometric, and topological information into a structured, vectorized format for GIS remains a significant challenge. Existing methods often struggle with irregular building shapes, noisy segmentation masks, and preserving sharp corners, leading to models that are either too dense (e.g., meshes, point clouds) or lack geometric regularity. This paper aims to bridge this gap by developing a method that regularizes building contours and generates compact, vectorized 3D models (Level of Detail 0 and 1) from readily available orthophotos and DSMs.
Methods
The proposed method consists of four main stages:
- Dominant Directions Extraction: Building contours are first extracted from segmented orthophotos. The normals of these contour edges are then smoothed using a bilateral filter that considers both angle similarity and geodesic distance. This smoothed data is fed into a RANSAC algorithm to detect dominant directions present in the building's outline.
- Target Alignment Direction Assignment: A Markov Random Field (MRF) formulation is used to assign each contour edge a target dominant direction. This step leverages a graph-cut approach to ensure that only the most significant directions are chosen, effectively regularizing the contour.
- Contour Deformation: A novel deformation energy is defined on a Delaunay triangulation of the contour polygon. This energy combines an alignment term (ensuring edges align with their target directions) and a distortion term (using the Advanced Most-Isometric ParameterizationS - AMIPS - to minimize geometric distortion). Minimizing this energy deforms the contour vertices to achieve alignment while preserving local geometry.
- Model Generation: The deformed contour is then simplified by retaining only corner vertices (those with large angles between adjacent edges). Connecting these vertices forms the Level of Detail 0 (LOD0) vectorized polygon model. The Level of Detail 1 (LOD1) model is generated by extruding the LOD0 contour to the average building height derived from the DSM.
Datasets
The authors evaluate their method on the public ISPRS 2D semantic labeling contest dataset (specifically the Vaihingen dataset), which includes high-resolution orthophotos and corresponding DSMs. They also use their own real-world orthophoto data for a specific experiment demonstrating the alignment deformation.
Key Insights and Ideas
- Contour Deformation as a Regularization Tool: The core idea is that by defining a deformation energy that penalizes deviations from dominant directions, irregular contours can be systematically regularized. This is a more flexible approach than simple simplification algorithms.
- Combining Global and Local Constraints: The method effectively combines global directional information (detected dominant directions) with local geometric constraints (AMIPS for minimal distortion) to achieve robust regularization.
- MRF for Direction Assignment: Using MRF for assigning target directions to edges is crucial for ensuring that the regularization process is consistent across the entire contour and favors globally dominant orientations.
- Preserving Sharp Corners: A key advantage highlighted is the ability to straighten edges while retaining sharp corners, which is often a challenge for traditional simplification algorithms.
Key Results
- The proposed method achieves high Intersection over Union (IoU) scores in contour quality comparison, outperforming methods like RDP and Poullis et al. [8] and Zhu et al. [21]. For instance, one comparison shows an IoU of 0.96 for their LOD0 model versus 0.94 for RDP and 0.87 for Zhu et al. [21].
- The method successfully generates regular and compact building models (LOD0 and LOD1) that adhere to CityGML standards.
- Qualitative results demonstrate the ability to recover the position of corner vertices accurately and reconstruct orthogonal neighboring edges correctly, even from noisy segmentation masks.
- The deformation energy minimization process shows a rapid drop in energy in the initial iterations, indicating efficient convergence.
Limitations and Future Work
- Focus on Buildings: The current work focuses solely on building modeling, excluding other urban elements like roads and vegetation.
- Manhattan Assumption: While the method doesn't strictly enforce the Manhattan assumption, its effectiveness is implicitly tied to the presence of dominant orthogonal or parallel directions in building outlines. Scenes lacking such regularities might pose challenges.
- Future Work: The authors suggest extending the method to include more common regularity constraints (e.g., parallelism, collinearity) and generating higher Levels of Detail (LODs) for the urban models.
Summary updated