Comments

Things to improve

• Would like to try out the Hough algorithm to see how it compares in customisation, robustness to outliers
• Fix the area estimation. Possibilities:
  • pcd triangulation for a 2d polygon in 3d space
  • Using vertical planes to clip roof planes
• Planes that are on similar enough should be grouped if the are within a certain area (saw the algorithm split one flat section into two). Probably is a way to avoid this happening in the core algorithm though

General thoughts

• Tweaking the figures on the algorithm to match the specific use case is important. i.e. Ransac threshold to within the flatness of a roof + lidar error margins.
• Preprocessing the point cloud seems like a very powerful tool. Gives the main algorithm the best chance to shine

Notes

Formats

• PLY - Polygon file format - format to store 3D data
• OBJ - format to store 3D data, supports lines/trimeshes geometry, materials, textures
• gltf/glb - supports animation, scene graph, textures

Python libraries

• Open3d - library for 3D data processing
• Trimesh - library for maniplating tri based meshes
• PDAL - library for manipulating point cloud data

Cleaning algorithms

• KD-Tree statiscical outlier removal (https://www.mdpi.com/1424-8220/21/11/3703/htm)

Plane fitting algorithms:

• least squares
• RANSAC
• Hough
• Principal component analysis

Key terms:

• Singular value decomposition
• Covariance matrix

Refs:

• https://stackoverflow.com/questions/35726134/3d-plane-fitting-algorithms
• https://stackoverflow.com/questions/28731442/detecting-set-of-planes-from-point-cloud
• https://stackoverflow.com/questions/38754668/plane-fitting-in-a-3d-point-cloud