Automatic Normal Orientation in Point Clouds of Building Interiors

Sebastian Ochmann, Reinhard Klein

Determining the orientation of surface normals in 3D point clouds is a fundamental challenge in geometry processing, particularly for complex indoor environments. While calculating the direction of a normal is straightforward via Principal Component Analysis (PCA), orienting it correctly—ensuring it points "outward" or "inward" consistently—is notoriously difficult in unstructured data.

Most existing methods rely on the assumption of a "closed 2-manifold" (a watertight solid), which real-world indoor scans rarely satisfy due to clutter, holes, and non-manifold topologies like shared walls. This paper presents a robust, automated pipeline that combines plane detection with multi-bounce path tracing to achieve high-accuracy normal orientation without requiring watertight geometry.

Methodology: Multi-Bounce Path Tracing

The proposed method transitions from raw points to higher-level geometric "patches" to improve both robustness and efficiency. The pipeline consists of several key stages:

  1. Plane Detection and Patch Generation: Using RANSAC, the algorithm identifies dominant planes and groups points into patches using a 2D occupancy grid. This reduces the problem size from millions of points to thousands of patches, drastically lowering computational costs.
  2. Orientation via Path Tracing (Phase 1): For each patch, the algorithm casts 50 rays into both hemispheres. Unlike traditional visibility methods that only consider the first hit, this approach uses multi-bounce path tracing (up to 8 bounces). Rays trapped in a room will bounce frequently and travel long distances, while rays pointing toward the exterior or into wall volumes will exit or terminate quickly.
  3. Refinement and Façade Correction (Phase 2): A voting mechanism ensures global consistency for patches on the same plane. Additionally, a second ray-casting pass identifies façade patches—often misoriented in indoor scans that "see" through windows—and corrects them.
Visual comparison of orientation results before (a) and after (b) refinement phases.

Evaluation and Key Results

The algorithm was evaluated on four large-scale, real-world datasets containing between 5.1 and 34.9 million points. Ground truth was established using known scanner positions from the original data acquisition.

The multi-bounce logic proved highly effective at distinguishing between interior spaces, wall volumes, and exterior clutter. By using the logarithm of path lengths, the algorithm prevents outlier rays (e.g., those escaping through windows) from skewing the statistical evidence provided by trapped rays.

Quantitative and qualitative evaluation across four large-scale indoor datasets.

As shown in the evaluation data, the pipeline achieved a final correctness rate between 98.68% and 99.42% across all datasets. While the initial plane detection step is the primary computational bottleneck (taking up to 4 minutes), the core orientation logic is remarkably efficient, consistently finishing in under 10 seconds even for the largest 35-million-point dataset.

Discussion and Limitations

The primary strength of this approach is its ability to handle "nested" topologies and non-watertight scans that cause traditional propagation-based algorithms to fail. However, it remains dependent on the initial detection of planar surfaces. Highly curved or organic architectural features may not be well-represented by the current patch-based logic. Future work could involve exploring faster patch generation techniques or alternative shape detection algorithms to further optimize the end-to-end runtime and expand the method's applicability to non-planar geometry.

Summary updated

BibTeX
@inproceedings{ochmann2019automatic,
  author = {Sebastian Ochmann and Reinhard Klein},
  title = {{Automatic Normal Orientation in Point Clouds of Building Interiors}},
  booktitle = {CGI},
  year = {2019},
  doi = {10.1007/978-3-030-22514-8\_55},
}
Figure