High-Resolution Plant Shape Measurements from Multi-view Stereo Reconstruction

Accurate high-resolution 3D models are essential for a noninvasive analysis of phenotypic characteristics of plants. Leaf surface areas, fruit volumes and leaf inclination angles are typically of interest. This work presents a globally optimal 3D geometry reconstruction method that is specialized to high-resolutions and is thus suitable to reconstruct thin structures typically occuring in the geometry of plants. Volumetric 3D models are computed in a convex optimization framework from a set of RGB input images depicting the plant from different view points. The method uses the memory and run-time efficient octree data structure for fast computations of high-resolution 3D models. Results show accurate 3D reconstructions of barley, while an increase in resolution of a factor of up to 2000 is achieved in comparison to the use of a uniform voxel based data structure, making the choice of data structure crucial for feasible resolutions.