RootModeling

Robust image-based 3D Modeling of Root Architecture

Abstract

Root system architecture (RSA) plays an important role in plant development and survival. The ability to accurately model and quantify properties of root architecture is fundamental for sustainability studies, crop improvement, and studies of plant-microbal interactions. Existing methods to model RSA either require a dense set of images or rely on 3D scanning methods for dense reconstruction. In this paper, we present an image-based technique for recovering complex 3D root geometry from a sparse set of viewpoints. Our solution incorporates 2D/3D root system topology as shape prior into the geometric reconstruction process. For every input view, we apply image matting for segmenting the root from the background. We then recover a 2D skeleton graph of the root from its matte image and find its corresponding 2D topology tree from the skeleton graph. Next, we present an iterative algorithm for computing the 3D topology tree that is most consistent with the set of 2D topology trees. Finally, we apply volumetric reconstruction for recovering the complete 3D root model from its 3D topology tree. We demonstrate our framework on roots of rice (Oryza sativa) and brachypodium (B. distachyon) and our experiments show that our method is robust and accurate.

Acquisition System

a single CCD-camera , Canon EOS Digital Rebel Xti SLR camera with 55 mm focal length and a resolution of 2592x3888 pixels.
a turntable used to obtain multiple views of the roots, with 20cm of diameter, its position can be specified with an accuracy of 1:00.
a calibration pattern used to extract point correspondence and run structure-from-motion on them to recover camera parameters.
a blue-light lamp used to illuminate the root during acquisition.

System Pipeline

1. Robust Skeletonization

To extract the skeleton graph from the cell complex, we use the robust thinning algorithm. This thinning algorithm formulated a skeleton significance measure, called medial persistence. Guided by this measure, the previous algorithm produces a family of topology and shape preserving skeletons whose shapes and composition can be flexibly controlled by the desired level of medial persistence.

2. 2D Topology Tree Construction

We first convert the skeleton graph into an undirected graph where the nodes are defined by leaves and junctions in the skeleton graph. We use the connectivity of the points in the skeleton to identify the leaf and junction nodes. Leaf nodes are detected by finding the endpoints, while the junction nodes correspond to the 3-connected and 4-connected points in the skeleton. We then find the edges that connect the nodes in the undirected graph by removing the interior points in the path between every pair of connected nodes.

3. 3D Root Topology Recovery

We next proceed to compute the 3D topology tree that is most consistent with the set of 2D topology trees. The core of our approach is to match a pair of trees. We developed an iterative tree matching technique that estimates the optimum match between two skeleton trees. We use a simple but efficient dynamic pruning technique that only discards unfeasible solutions, rather than discarding solutions based on their cost.

4. Model Recovery

Once we recover the optimum 3D topology tree, we apply a revised volumetric reconstruction approach to recover the complete model of the root. For each edge in the 3D topology tree, our approach applies a space carving method on a local volume to estimate the branch geometry (e.g., shape and size). Our solution is based on the observation that topology tree casts important geometry constraints.

Results

A gallery of models recovered by our algorithm