A deep learning-based VMSNet for high-precision point cloud segmentation and ripe tomato diameter phenotyping in greenhouses

Abstract

Ripe tomato fruit display diverse 3D morphologies driven by genetics, environment, and management, yet these differences remain hard to quantify in the absence of precise point-cloud segmentation tools. This paper proposes the VMSNet to accurately segment tomato fruits and extract phenotypic traits based on the segmentation results, including horizontal and vertical diameters. Point clouds are obtained through depth cameras. After preprocessing and labeling the fruits, a dataset is established using global enhancement and local enhancement. On the base framework of PointNet++, the downsampling method was replaced, a multi-scale attention module (MS_A) was integrated, the combination scheduling strategy was optimized, and VMSNet was constructed. Following segmentation, the fruit growth direction is estimated by density-weighted method, and principal component analysis (PCA) is used to establish a rotation plane. By rotating according to the slicing angle, the fruit point cloud is completed and fitted into an ellipsoid. Random Sample Consensus (RANSAC) is used to smooth the outliers. The OBB is applied to extract the horizontal and vertical diameters, which are compared with measurement to verify the algorithm’s accuracy. The results indicate that the accuracy of VMSNet in segmenting ripe tomato fruits is 97.96%. The correlation coefficients R 2 between the calculated and measured values of the horizontal and vertical diameters reached 0.89 and 0.86, respectively. This proposed proposal provides robust point cloud segmentation and completion for phenotypic analysis for other same species greenhouse crop.

Plant phenotyping relevance

深度学習による点群分割とトマト果実径の抽出手法を開発し、実測値との比較で精度検証しており、植物表現型取得が研究の中心である。

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.