Data driven tree modeling

Abstract

We propose methods to model trees using minimal and maximal data. For the minimal data, we could model a tree given its segmentation only. The non-parametric synthesis is improved in terms of its run-time efficiency. With more data such as its image and initial skeleton, we could synthesis complete branches and generate realistic leaves according to the image. For the maximal data, we extend the framework to work on laser scanned points and multiple images, in particular the Google R5 data. With the laser data, we could automatically segment the tree points and infer their initial skeletons. The non-parametric synthesis is extended to work on the sparse laser points and noisy 2D segmentations. We further study the simplification of tree models. We propose a method to simplify lateral tree by topological reduction. For mesh simplification, invisible branches are removed by ray-tracing. It is further simplified by reducing the interpolations of both curves and surface sub-divisions. For leaves simplification, we propose a billboard cloud approach that let all the leaves to fit to the well oriented billboard cloud, which does not require any rendering technique for it to work. Our method is evaluated by a variety of trees and by a street of trees in San Francisco.

Date of Award	2011
Original language	English
Awarding Institution	The Hong Kong University of Science and Technology