Constraint-based automated reconstruction of grape bunches from 3D range data for high-throughput phenotyping

For grape bunches, a loose architecture is one of the most important physical barriers to avoid fungal infections, making the phenotyping of grape bunches, i.e., the derivation of the morphological attributes, a major goal for breeders. The stem skeleton plays an important role in determining the architecture, but is for grape bunches with berries ready for harvest usually completely occluded by the berries, making it invisible in the data. We introduce geometrical and topological constraints to encode knowledge about geometry and topology of the plant organs to foster efficient sampling of infinitely large hypotheses spaces for target objects with a high amount of self-occlusion and high within-class-scatter. To further support the reconstruction process, we introduce a new variant of the Reversible-Jump-Markov-Chain-Monte-Carlo algorithm with locally optimised jumps, meaning that after using a jump to select a model, a parameter optimisation is performed on the changed parts. We show the usability of our method by deriving established phenotypes, like berry diameters and average pedicle length, from scans created in a lab environment, yielding promising Pearson correlation coefficients between 0.7 and 0.9 on 52 grape bunches spread over four grapevine cultivars. Our approach reduces the active working time required for grape phenotyping by the factor of 12 compared to the manual method currently used at the Institute for Grapevine Breeding Geilweilerhof in Siebeldingen. This speed-up can be increased even more by taking scans in the field. On this application, we achieve first results that are for symmetrical grape bunches comparable to lab scans.