GrapeCPNet: A self-supervised point cloud completion network for 3D phenotyping of grape bunches

Abstract

The measurement of phenotypic parameters of fresh grapes, especially at the individual berry level, is critical for yield estimation and quality control. Currently, these measurements are done by humans, making it costly, labor-intensive, and often inaccurate. Advances in 3D reconstruction and point cloud analysis allow extraction of detailed traits for grapes, yet current methods struggle incomplete point clouds due to occlusion. This study presents a novel deep-learning-based phenotyping pipeline designed specifically for 3D point cloud data. First, individual berries are segmented from the grape bunch using the SoftGroup deep learning network. Next, a self-supervised point cloud completion network, termed GrapeCPNet, addresses occlusions by completing missing areas. Finally, morphological analyses are applied to extract berry radius and volumes. Validation on a dataset of four fresh grape varieties yielded $R^2$ values of 85.5% for berry radius and 96.9% for berry volume, respectively. These results demonstrate the potential of the proposed method for rapid and practical extraction of 3D phenotypic traits in grape cultivation.