Maize stem–leaf segmentation framework based on deformable point clouds

Abstract

The efficacy of three-dimensional (3D) point clouds in studying crop morphological structures is based on their direct and accurate data presentation ability. With deep-learning integration, organ segmentation from point clouds could serve as the basis for tremendous advancements in organ-level phenotyping. However, despite the potential, the acquisition of a sufficient number of annotated plant point clouds for practical model training remains a major hurdle. To help overcome this limitation, we constructed a 3D point-cloud dataset specifically for maize stem–leaf segmentation encompassing 428 maize plants ranging from 2 to 12 leaves. We also developed a point cloud enhancement strategy that uses highly controllable deformations to improve the morphological diversity of the training set significantly, while preserving the local geometric features of organs. Our dataset supports the generation of abundant training data from a limited number of labelled data, and we also provide a segmentation framework based on the augmented data to validate the efficiency of our enhancement technique. Two labelled data items were randomly chosen from our plant dataset based on every leaf number, yielding 22 labelled data items total, to produce several deformed point clouds for training the PointNet++ semantic segmentation model, as well as the hierarchical aggregation for the 3D instant segmentation (HAIS) model. These models were tested on 406 datasets, where the PointNet++ model secured a 91.93 % mean intersection-over-union (mIoU) in semantic segmentation and the HAIS model obtained an 89.57 % mean average precision (mAP) in instance segmentation. Following post-processing, an instance segmentation result of 93.74 % mAP was achieved with the HAIS model. These findings demonstrate that our method allows for the efficient training of organ segmentation models with minimal labelled data input in a reduced timeframe. Moreover, it offers an effective tool for point-cloud parsing in maize phenotyping research. Our Maize dataset is available from https://github.com/syau-miao/SignleMaizePointCloudDataSet.git, and the source code of our method can be found at https://github.com/yangxin6/Deformation3D.git.