VGDS-PointNet++ for organ segmentation and phenotypic trait estimation in greenhouse tomato seedlings.

Abstract

Introduction To accurately segment point clouds and quickly calculate leaf length and stem diameter, thereby enabling phenotypic analysis and variety selection of greenhouse tomato plants, this paper proposes a voxel grid downsampling (VGDS)-PointNet++-based model for point cloud segmentation and trait calculation. Methods The point clouds of the tomato canopy were acquired using a depth camera. After labeling, point cloud augmentation was performed, and the tomato point cloud dataset (TPCD) containing 1,552 sets of data was rebuilt. Voxel grid downsampling was applied to replace the original sampling strategy of PointNet++. Models of PointNet, PointNet++, VGDS-PointNet++, and Point Transformer were trained with the TPCD and compared on segmentation quality with accuracy and mean Intersection over Union (mIoU). After segmentation, skeletal morphology was fitted for non-occluded leaves by applying a series of surface fitting techniques. The leaf lengths and stem diameters were automatically calculated and compared with the manually measured values. Results The validation results showed that the average runtime of voxel grid downsampling was 0.132 s, which was lower than under the same number of sampled points. Compared to the other three models, the proposed model had higher accuracy and mIoU, reaching up to 96.80% and 88.95%, respectively. The proposal's accuracy and mIoU increased by 3.9% and 4.45% over PointNet++, respectively. The determination coefficient R 2 between the automatic calculation and manual measurement values of leaf length and stem diameter was 0.93 and 0.87, respectively. Discussion This can help extract phenotypic traits of tomatoes using depth cameras.

Plant phenotyping relevance

深度カメラ点群のセグメンテーションと葉長・茎径の自動推定手法を開発し、手測定および複数モデルと比較検証しており、表現型取得が研究の中心である。

Code and data availability

The supplied blocks describe a tomato point cloud dataset (TPCD) and VGDS-PointNet++ model, but contain no data availability statement, no public repository deposit, and no authors' URL for dataset, code, or trained models. The Data availability statement section is listed in the outline but its text is not present in,

No evidence-backed public reproduction asset is currently recorded.