Object Reconstruction

You arrive here with a mesh. Chapter 15 reconstructed it from whatever the points described, in one vocabulary for a building, a tree, and a power-line span alike, and it records where the surface lies to centimetres. That is the whole of what it records, since a triangle carries no name. So you can measure anything on that mesh and report nothing from it. Yet the cadastre asks you for the roof form and the eave height, the forest inventory for the stem volume, the utility for the ground clearance. Each of those names a part, an eave, a stem, a conductor, that a mesh does not distinguish from any other patch of surface. This chapter fits a different model to each class of object, one that carries those names.

How much does the model matter against the measurement behind it? Here is a number from later in the chapter. Take one street tree, scanned once with a handheld mobile scanner, and reconstruct it with two published pipelines. AdTree returns 3.47 m³ of wood and arbor returns 0.82 m³, a factor of 4.2 from the same returns (figure 16.9). At breast height the two agree to within a few centimetres, because the stem is well sampled and both read the same measurement there. But the crown is not well sampled, and there the model supplies what the returns do not. So a reconstructed object is a measurement and a set of assumptions in one file, and knowing which is which is your job in every section that follows.

Four questions run through the domains that follow. Where do you assume structure, in the parts or in the whole? Which primitive does the object admit, a plane, a cylinder, a catenary? What did the survey constrain, and what arrived from a catalogue? And what may you claim when you hand the model over?

Where does the input come from? Not from that mesh. You work from the classified and segmented point cloud. Here chapter 12 supplies the per-class points. Here chapter 11 isolates the individual objects. And chapter 9 anchors heights and footprints. The eigenvalue features of chapter 10 flag the planar and linear structures the fitting stages exploit. But none of those stages marks a facet as a roof plane or an edge as its ridge, which is what this chapter adds. Whatever the domain, you follow the four stages of figure 16.1, classify the points, isolate the object, fit the model, export the structured result. What comes out feeds the city-scale, forestry and infrastructure workflows of chapter 18. Let's start with what a structured model has to contain, since that settles how much of the rest you need.