Automatic 3D Plant Organ Instance Segmentation Method Based on PointNeXt and Quickshift+.

Abstract

Organ instance segmentation of 3D plant point clouds is a crucial prerequisite for organ-level phenotype estimation. However, most current cloud segmentation methods are usually designed for specific crop, hardly fit for both monocotyledonous and dicotyledonous crops which have significant structural differences. This study therefore proposed a two-stage method with higher generalization ability for single-plant organ instance segmentation based on PointNeXt and Quickshift++. The effectiveness of this method was tested on different types of crops. The dataset includes point clouds of 122 self-acquired sugarcanes, 49 open-accessed maizes, and 77 open-accessed tomatoes. The improved PointNeXt model was trained to implement the semantic segmentation of stems and leaves. The average mOA and mIoU on the test set reaches 96.96 ​% and 87.15 ​%, respectively. The Quickshift++ algorithm was then applied to encode the global spatial structure and local connections of plants for rapid localization and segmentation of leaf instance. Our approach outperformed four SOTA methods, ASIS, JSNet, DFSP, and PSegNet in terms of both quantitative and qualitative segmentation results, achieving average values for mPrec, mRec, mF1, and mIoU of 93.32 ​%, 85.60 ​%, 87.94 ​%, and 81.46 ​%, respectively. The proposed method also yields excellent results for several other plants in their early stages, indicating its generalization ability and applicability for organ instance segmentation for different plants, thus providing a powerful tool for plant phenotypic research.

Plant phenotyping relevance

植物器官の3D点群から茎・葉のインスタンスを抽出する手法を開発・比較検証しており、器官レベル表現型推定のための中心的なフェノタイピング手法である。

Code and data availability

The authors state their dataset and code were uploaded to a public GitHub repository, and the paper's maize/tomato point cloud inputs come from the public Pheno4D dataset. Both are paper-specific, public, and actionable.