Geometric modeling and shape analysis of 3D point clouds

This thesis investigates the complex problem of automatically reconstruct large-scale scenes from 3D point clouds. The reconstruction problem is decomposed into two major components, namely, primitives and parts. While primitives are geometric shapes (e.g., planes, cylinders, etc.) and their connections (e.g., joints), parts are relatively isolated objects that can be hardly interpreted by geometric shapes. Two different strategies are presented to handle these two reconstruction problems. ❧ In primitive reconstruction, two systems focusing on different targets and scenes are presented. The first one reconstructs pipe-runs from industrial site point clouds, because they are dominant structures in most industrial sites. The key idea is that by adopting statistical analysis over point normals, global similarities are discovered from raw data to guide primitive fitting, thus increasing robustness to data noise and incompleteness. The second one extracts pole-like objects from urban point clouds and posed images. Pole-like objects are common in urban scenes but are difficult to model because their thin structures are usually undersampled in images and point clouds. The presented method takes advantage of the complementary information from 3D point clouds and 2D posed images to recover these objects. In both systems, the resulting model is more than a collection of 3D triangles, as it contains semantic labels for primitive shapes. ❧ In part reconstruction, a modeling-by-recognition strategy is followed. Instead of directly working on a noisy scan, a matching template point cloud is retrieved from a part library. The library object and input object should match in basic category, but can be dissimilar in geometry and topology. Then, geometric analysis is applied on template and input point cloud to accomplish two tasks. The first one is to automatically compute dense correspondences between input and template scans, thus making it possible to transfer real-world color to template models. The method segments both point clouds into parts and then computes part-level correspondences between source and target scans. The dense mapping allows color or other parameter transfers. The second one is to segment input scans using a small set of pre-segmented template point clouds as examples. The main idea is to register the target shape with exemplar shapes in a piece-wise rigid manner, so that pieces under the same rigid transformation are more likely to be in the same segment. The transferred segmentation of real scans is a key step towards shape understanding, and a good initialization of part-based dense correspondences.