Multi-PointNet++: A Multi-Scale Local Interaction Network for Plant Point Cloud Segmentation

Abstract

To address the challenges of complex structures, large organ-scale variations, and dense internal organization in plant 3D point clouds, This paper proposes Multi-PointNet++ algorithm based on multi-scale local interaction to enhance the accuracy and detail extraction capability of plant point cloud segmentation. We first construct the EMA-Shuffle attention module during feature extraction, combining Enhanced Multiscale Attention (EMA) with Shuffle Attention to enable efficient local feature interaction through multi-scale fusion. At the same time, the OREPA reparameterized convolution module enhances feature extraction and model performance while keeping the structure lightweight. For feature fusion, we design a pyramid fusion module based on SFPN that integrates triple sampling strategies and bidirectional feature propagation to effectively align geometric information across scales. Our method improves mIoU by 7.1% and 5.3%, and mAcc by 6.7% and 13.7% on the public PLANesT-3D and self-built sunflower datasets, respectively, significantly outperforming other models. This approach offers robust 3D semantic support for high-throughput plant phenotype extraction.

Plant phenotyping relevance

植物3D点群からの器官・構造抽出を目的とする新規セグメンテーション手法を開発し、公開データセットと自作データセットで性能検証しているため、植物表現型取得法が中心である。

Code and data availability

公開論文であることは確認できましたが、現在の公式API・許可済み取得経路では本文を自動取得できませんでした。

No evidence-backed public reproduction asset is currently recorded.