Semantic labeling and reconstruction of grape bunches from 3D range data using a new RGB-D feature descriptor

Abstract

In the context of grapevine breeding, high precision and automated phenotyping plays an important role in order to screen breeding material (e.g. seedlings) or to characterize genetic repositories with high-throughput. Grape bunches hereby reveal a large variability regarding size, shape, compactness and color. We design and evaluate a new RGB-D descriptor for the semantic labeling of grape bunches. For this, we examine RGB and HSI color spaces and combine them with Fast-Point-Feature Histograms. With the best combination of FPFHs and the hue channel we achieve an average F-value of 88.61%, outperforming classical descriptors like PFHRGB and SHOTColor by at least 8%. Additionally, we show a new method for the derivation of parametric reconstructions of the elliptical berries based on a least squares fitting, yielding Pearson correlation coefficients of 0.8 and 0.9 for the main diameters of the berries.