Curve-based Stereo Matching for 3D Modeling of Plants

Realistic image-based 3D reconstruction of narrow plants, such as wheat and barley, is challenging because of plant’s lack of a well defined surface, complex self-occlusions, and high degree of self-similarities. In this paper we present a new approach for image-based modeling of plants using a pair of cameras in a stereo setup. Our key idea is that while color and texture features are unlikely to match across images, local leaf shape, in particular the shape of its boundary curves, will. We propose a stereo matching algorithm that uses boundary curve features to build a dense depth map. We show that the local orientation of boundary curves can be used to recover the 2D leaf structure of plants, which in turn can be used for matching plant leaves across multiple views in a narrow baseline stereo setup. The proposed method is robust to slight movements of the leaves during imaging, which is an important concern in phenotyping narrow and fragile-leaf plants. We demonstrate results on a number of plants with varying complexity and discuss the performance and limitations of the proposed framework.