TomatoPGT: A 3D point cloud dataset of tomato plants for segmentation and plant-trait extraction

TL;DR

TomatoPGT is a 3D dataset of tomato plants that helps researchers study plant structure and traits without damaging the plants.

Contribution

TomatoPGT introduces a 3D point cloud dataset with detailed annotations and graph-based representations for plant phenotyping.

Findings

The dataset includes 42 scans of tomato plants with RGB images and dense colored point clouds.

Manual annotations and graph-based traits like internode length and angles are provided for research use.

Abstract

Three-dimensional (3D) point-cloud phenotyping enables non-destructive and repeatable characterization of plant architecture, supporting the measurement of traits such as internode length, branching topology, and organ orientation. This article presents TomatoPGT (Tomato Plant Graph Twin), a 3D tomato dataset designed for research on semantic/instance segmentation, graph-based structural representation, and graph-derived phenotypic trait extraction.

The dataset contains 42 scans from three greenhouse-grown tomato plants acquired across early to mid-vegetative development using a rotational multi-view imaging system. Each scan consists of 60–70 overlapping RGB images captured under uniform illumination and reconstructed into a metrically scaled dense colored point cloud using Structure-from-Motion and multi-view stereo. TomatoPGT provides: (i) multi-view RGB images, (ii) dense colored point clouds, (iii) manually curated semantic and instance annotations at organ level, (iv) graph representations encoding plant topology and geometry, and (v) tabulated phenotypic traits computed deterministically from the graphs (internode length, insertion angles, and phyllotactic angles). TomatoPGT supports reproducible development and evaluation of 3D phenotyping pipelines, including learning-based segmentation and graph-based modeling of plant architecture.

Figures

Temporal samples from the TomatoPGT dataset.Fig 1 dummy alt text

Schematic of the simplified structural annotation used in TomatoPGT.Fig 2 dummy alt textTable 1Simplified annotation schema and ontology for TomatoPGT.Table 1 dummy alt textSemantic ClassBiological Description*Common Annotation ErrorsRoot-Node*Origin / basal attachment of the plantMissing root assignment*Junction-Nodes*fork on the main stemConflation with stalk attachment*mainStem-Seg*Stem segments / internodesBroken stem chains; duplicate instance IDs*Compound Leaf-Node*Aggregated leaflets and rachisInconsistent leaflet grouping*Stalk-Seg*Petiole (connecting Junction to Leaf)Overlapping leaf/

Visualization of the generated graph-based representations.Fig 3 dummy alt textTable 3Graph table preview.Table 3 dummy alt textplant_idgraph_nodesgraph_edges#junctions#compound_leaves /Ct#suckersBBH_0513202565120BBH_0515202565120BBH_0520202576120BBH_0512202576120

Overview of the TomatoPGT data pipeline, illustrating image acquisition, 3D reconstruction, annotation, graph extraction, and phenotypic trait computation.Fig 4 dummy alt text

Two additional markers were treated as Independent Check Points (ICPs) for geometric validation D and E.Fig 5 dummy alt text

ICP error distribution.Fig 6 dummy alt textFig. 7ICP stability over time (error vs scan date).Fig 7 dummy alt text

Full text

Specifications TableSubjectBiologySpecific subject area*3D tomato plant phenotyping using point-cloud segmentation, graph-based structural representation, and graph-derived trait extraction.*Type of data

• −Raw images: JPEG (multi-view)-Dense point clouds: PLY (XYZ, RGB; normals)
• −Annotated point clouds: TXT (XYZ, RGB, class_id, instance_id, annotation RGB)
• −Graph representations: JSON (nodes, edges, geometry, metadata)
• −Phenotypic traits: CSV (graph-derived measurements) *Note:Due to repository storage constraints, full RGB image sets are provided for Celebrity Hybrid and Beefsteak, while Big Boy Hybrid is released with point clouds, annotations, graphs, and traits. Big Boy Hybrid RGB images are provided in the Zenodo archive.Data collectionData were acquired using a rotational multi-view imaging system with a motorized turntable and synchronized camera triggering. A Sony A6000 (24 MP;16mm lens) captured 60–70 overlapping images per scan at 5–10° increments under uniform illumination inside an LED lightbox. Reconstructions were generated using Agisoft Metashape[1]. with consistent settings. Point clouds were filtered to remove background/noise before manual semantic and instance annotation and graph extraction. Reconstructions were metrically scaled using a printed marker sheet; no destructive manual plant measurements were collected, and all traits are derived from the reconstructed geometry and graph representation.Data source locationPlants were grown under rain shelter with controlled fertigation at the Texas A&M AgriLife Research, Corpus Christi, TX, USA and controlled indoor environment at the Texas A&M University–Corpus Christi Agricultural Robotics Lab, USA (latitude: 27.7136 ° N, longitude: −97.3256 ° W)Data accessibilityRepository name 1: Mendeley[2].Data identification number: DOI: 10.17632/72md54c7n7.1Direct URL to data: https://data.mendeley.com/datasets/72md54c7n7/1For software and supplementary materialsRepository name 2: Zenodo[3].Data identification number: DOI 10.5281/zenodo.18651881Direct URL to data: https://doi.org/10.5281/zenodo.18651881Related research articleP. Nethala, D. Um, K. Lee, M. Bhandari, O. F. Montero, and N. Vemula, “Techniques for plant feature extraction and segmentation: A survey,” Remote Sensing, vol. 16, no. 23, 2024. Available: https://doi.org/10.3390/rs16234370[4].

Value of The Data

1

• •Plant phenotyping is the quantitative study of observable traits that describe plant physiology and growth. While digital phenotyping methods have largely relied on 2D imagery, plants exhibit inherently 3D structural and spatial features, such as internode length, branching, and leaf angles, that directly affect light interception, photosynthesis, and yield potential. Hence, accurate 3D phenotyping is essential to fully capture these spatial characteristics.
• •We introduce TomatoPGT, a comprehensive 3D tomato plant graph dataset that includes annotated point clouds, semantic and instance labels, semantic graph-based representations, and phenotype trait extractions.
• •Two custom software tools complement the dataset: Cloud-Seg, an interactive labelling platform for 3D point-cloud annotation, and Cloud-Graph, an algorithmic graph-generation tool. These tools support standardized workflows for segmentation, graph topology, and trait extraction, enabling reproducible evaluation of phenotyping algorithms.
• •Graph-based representations preserve plant topology and geometry while reducing data complexity, enabling efficient downstream analysis.
• •The dataset supports research in semantic segmentation, graph learning, phenotypic trait extraction, and digital plant modelling.

DatasetCrop(s)Annotations (publicly available)Notes / ReferenceROSE-XRoseSemantic organ labelsX-ray CT + point cloud; organ-level annotations [5].Soybean-MVSSoybeanSemantic102 plants across growth stages; MVS reconstruction [6].PLANesT-3DPepper, Rose, RibesSemantic + Instance34 real plants; leaf/stem + leaflet instances [7].Pheno4DMaize, TomatoSemantic + Instance + Time-seriesRegistered temporal point clouds [8].LAST-StrawStrawberrySemantic + Instance + Time-series + Skeletons84 scans of 6 plants; rich trait benchmarks [9].TomatoWURTomatoSemantic + Instance + Connected skeletons + Manual traits44 plants; shape-from-silhouette dataset [10].Crops3DMaize, Cabbage, Potato, Cotton, Rapeseed, Rice, Tomato, WheatSemantic ± Instance1230 colored 3D point clouds; multi-sensor collection [11].MaizeField3DMaizeSemantic + Instance520 annotated field-grown plants; TLS and NURBS models [12].TomatoPGT *(this work)*TomatoSemantic + Instance + Semantic Graph representations+ Phenotypic Trait extraction+ Software tools42 Scans; Camera rotational imaging system [2]; includes Cloud Graph and Cloud-Seg software[3].

Background

2

Accurate plant phenotyping is essential for understanding the complex interaction between genotype (G), environment (E), and management (M), commonly expressed as G × E × M [10,13,14]. Many phenotypic traits relevant to plant growth and productivity such as branching architecture, internode length, and organ orientation are inherently three-dimensional, motivating the use of 3D representations for structural plant analysis [15].

Recent advances in photogrammetry have enabled the widespread use of 3D point clouds for plant phenotyping. Several public datasets now provide semantic and instance-level annotations of plant organs from reconstructed point clouds, including ROSE-X [5], Soybean-MVS [6], Pheno4D [8], and PLANesT-3D [7]. While these datasets have significantly advanced organ-level segmentation, they generally focus on point-wise labeling and do not provide explicit structural representations of plant architecture or associated phenotypic trait measurements within a unified framework.

Graph-based representations offer a compact and interpretable abstraction of plant structure by encoding nodes (e.g., junctions and organ attachment points) and edges (e.g., stem and stalk segments). Such representations are well suited for modeling plant topology, supporting structural analysis, and enabling the computation of phenotypic traits that depend on both geometry and connectivity.

To support research along these lines, we introduce TomatoPGT (Tomato Plant Graph Twin Dataset), which integrates annotated 3D point clouds with explicit graph-based structural representations and derived phenotypic traits. The dataset builds upon prior 3D plant datasets by providing, for each plant, semantic and instance annotations, graph representations encoding plant topology, and tabulated phenotypic measurements derived consistently from those graphs.

Together with the accompanying Cloud-Seg and Cloud-Graph tools, TomatoPGT provides a unified resource for studying segmentation, graph-based plant modeling, and trait extraction from 3D data. The dataset is intended to facilitate reproducible experimentation across multiple stages of the 3D plant phenotyping pipeline and to support the development and evaluation of methods that operate beyond point-wise representations.

Data Description

3

Overall dataset structure

3.1

The TomatoPGT dataset is organized at the plant level. Each plant folder contains temporally ordered scans representing repeated acquisitions during vegetative development. The dataset is distributed across two repositories: Mendeley Data (primary dataset) and Zenodo (software and supplementary metrics).

Mendeley Data Repository (Primary Dataset)

Each unzipped plant folder follows the structure:

Zenodo Repository (Software + Supplementary Material)

The Zenodo archive (DOI: 10.5281/zenodo.18651881) contains:

Data modalities and previews

3.2

1. Raw Images and Point Clouds: The RGB_IMAGES folder contains 67 overlapping multi-view photos per scan, serving as the raw input for 3D reconstruction. The reconstructed point clouds (_R_PC) are metrically scaled and stored with 3D coordinates, color, and normals (x, y, z, r, g, b, nx, ny, nz). The dataset captures the morphological progression of the plants. (Fig. 1) illustrates the translation of raw photogrammetric models into semantic point clouds and, subsequently, into topological graph representations.Fig. 1. Temporal samples from the TomatoPGT dataset.Fig 1 dummy alt text

2 Semantic and Instance Annotations: The Annotated_TXT files align index-wise with the point clouds. To prioritize architectural topology over highly variable leaflet details, a simplified structural annotation schema was applied (Fig. 2). The complete annotation ontology is detailed in Table 1, and the annotated point clouds are shown in second row (Annotated PC) of Fig. 1.Fig. 2. Schematic of the simplified structural annotation used in TomatoPGT.Fig 2 dummy alt textTable 1Simplified annotation schema and ontology for TomatoPGT.Table 1 dummy alt textSemantic Class****Biological DescriptionCommon Annotation Errors**Root-NodeOrigin / basal attachment of the plantMissing root assignmentJunction-Nodesfork on the main stemConflation with stalk attachmentmainStem-SegStem segments / internodesBroken stem chains; duplicate instance IDsCompound Leaf-NodeAggregated leaflets and rachisInconsistent leaflet groupingStalk-SegPetiole (connecting Junction to Leaf)Overlapping leaf/stem boundariesSucker-Seg**Sucker / Branch / Lateral vegetative shootsMisclassification as primary stems

A preview of the resulting annotated instances extracted per scan is provided in Table 2*.*Table 2. Preview of the annotation table.Table 2 dummy alt textplant_idann_points(Million)#junctions_inst#compound_leaves_inst#suckersBBH_051320250.033100BBH_051520250.025100BBH_052020250.055100BBH_051220250.142100

4. Graph-Based Representations: The Graphs folder provides semantic representations that approximate the plant’s structural topology (Fig. 3). These JSON files encode node lists (3D coordinates), edge lists (connections), and sampled centerline geometries. A preview of the topological complexity extracted across the dataset is provided in Table 3.Fig. 3. Visualization of the generated graph-based representations.Fig 3 dummy alt textTable 3Graph table preview.Table 3 dummy alt textplant_idgraph_nodesgraph_edges#junctions#compound_leaves /Ct#suckersBBH_0513202565120BBH_0515202565120BBH_0520202576120BBH_0512202576120

In Fig. 3. (Left) The generated topological plant graph. Distinct semantic components are color-coded: main stem internodes (blue), stalks (orange), lateral suckers (magenta), compound leaves/tips (green), cotyledons (cyan), and primordia (yellow). Nodes are represented as distinct spheres: root node (brown) and junction nodes (blue). Computed geometric traits, including insertion angle arcs (red) and phyllotaxy arcs (cyan), are explicitly modelled. (Right) The extracted graph topology anchored by sampled centerline points (red) successfully overlaying the corresponding semi-transparent 3D raw point cloud.

5. Phenotypic Traits: Computed directly from the graph geometry, traits including plant height, internode lengths, stalk insertion angles, and phyllotactic angles are provided as CSV files.

6. Code and documentation: CloudSeg and CloudGraph software tools, environment specifications, and example usage instructions are hosted on Zenodo[3]. The repository includes pre-built Windows wheels, Example datasets, Environment configuration files, A sample Agisoft Metashape project for reconstruction reference, Validation per scan report, Tabulated summary of the annotated instances per plants organ class.

Experimental Design, Materials and Methods

4

Data Generation Pipeline and Acquisition Setup A comprehensive data pipeline (Fig. 4*) was developed to facilitate image acquisition, 3D reconstruction, annotation, graph extraction, and phenotypic trait computation.*Fig. 4. Overview of the TomatoPGT data pipeline, illustrating image acquisition, 3D reconstruction, annotation, graph extraction, and phenotypic trait computation.Fig 4 dummy alt text

In this study, A controlled 3D acquisition system was developed to capture tomato plants for high-resolution reconstruction (Fig. 1, top row). A Sony A6000 (24 MP, 16 mm lens) was mounted on a fixed rig, while the potted plant was placed on a 360° motorized turntable (6 W, 110 V, 15 rpm). The turntable was rotated in 5–10° increments, producing 67 overlapping images per scan.

Image capture was synchronized using a custom Python script communicating with an Arduino microcontroller via serial control, enabling a “stop–capture–rotate” sequence to ensure motion stability and uniform overlap. All acquisitions were performed inside a 47″ × 39″ × 78″ LED lightbox to minimize shadows and illumination variation. A printed fiducial marker sheet [16] (0.032 m spacing) was placed beneath each plant to provide geometric reference for scaling and coordinate normalization. Although the point clouds are metrically scaled using a marker sheet, no independent manual measurements were collected for cross-validation. Therefore, all trait values in the dataset are derived solely from the 3D reconstructions and the skeleton geometry.

3D Reconstruction and Metric Multi-view images were processed using Agisoft Metashape Professional [1] following a standard Structure-from-Motion (SfM) workflow (Fig. 1) first row. Processing steps included high-accuracy image alignment, bundle adjustment, and dense point cloud generation using mild depth filtering, culminating in the export of colored PLY models.

Consistent with standard close-range SfM workflows for accurate metric reconstruction [17,18], no independent external camera pre-calibration (e.g., checkerboard calibration) was performed prior to scanning. Instead, camera intrinsics including focal length, principal point, and radial distortion coefficients were estimated automatically via self-calibration during the bundle adjustment phase. This joint optimization of camera intrinsic parameters and extrinsic poses is highly robust and ensures the lens distortion model reflects the exact in situ optical state at the time of capture [19].

Absolute metric scaling and orientation for the reconstructions were enforced strictly through a local coordinate system established by marker-based constraints. Specifically, three markers were defined as Ground Control Points (GCPs): A (0, 0, 0) served as the origin, B (0.032, 0, 0) established the metric scale and primary axis, and C (0.032, 0.032, 0) defined the XY ground plane (Fig. 5). The complete, optimized camera calibration matrices and resulting parameters for all scans are openly accessible in the associated Zenodo data repository.Fig. 5. Two additional markers were treated as Independent Check Points (ICPs) for geometric validation D and E.Fig 5 dummy alt text

Internal geometric validation: Because manual plant measurements can be destructive, the geometric fidelity of the reconstructions was assessed using an Independent Check Point (ICP) framework. While markers A, B, and C constrained the coordinate system, two additional markers (D and E) were withheld from the alignment phase and utilized strictly for validation.

For each scan i, the empirical Euclidean distance (D_e,i_) between a given pair of validation markers was computed from their reconstructed 3D coordinates. The geometric error (∈i) was defined as the difference between this empirical distance and the theoretical distance (D_t_) derived from the known physical marker geometry:

$$[eqn]$$

Across scans, reconstruction consistency was summarized using:

• •Mean error $[eqn]$
• •Standard deviation of $[eqn]$
• •Root Mean Square Error (RMSE)

$$[eqn]$$

Across all temporal scans N, reconstruction consistency was evaluated by calculating the mean error (bias), Mean Absolute Error (MAE), and Root Mean Square Error (RMSE) for three independent check pairs (B–D, B–E, and D–E). As reported in (Table 4), sub-millimeter RMSE and MAE values across all pairs (0.301 mm and 0.151 mm, respectively) indicate highly stable metric scaling and negligible spatial distortion.Table 4. Independent geometric validation.Table 4 dummy alt textValidation pair (ICP)Theory (mm)Mean error (mm)RMSE (mm)MAE (mm)B-D (Step Check)320.04370.20300.0988B-E (Diagonal Check)45.2548−0.00580.19910.1262D-E (Step Check)320.01600.30090.1505

Furthermore, the distribution of these ICP errors (Fig. 6) is cantered near zero, confirming that residual deviations represent random reconstruction noise rather than systematic scaling bias. Temporal analysis of ICP errors across scan dates (Fig. 7) demonstrates that metric accuracy remained perfectly stable over time, exhibiting no degradation as canopy complexity increased during mid-vegetative development. Because all downstream graph extraction and phenotypic trait computations (e.g., node-to-node distances and vector-based angle computations) are derived directly from these point clouds, this internal validation explicitly establishes the geometric reliability and scale-invariance of the entire analytical pipeline.Fig. 6ICP error distribution.Fig 6 dummy alt textFig. 7ICP stability over time (error vs scan date).Fig 7 dummy alt text

Point-Cloud Preprocessing and Annotation Pipeline Raw point clouds underwent initial preprocessing to isolate the plant structure using mask-based filtering and statistical outlier removal. Semantic and instance annotations were subsequently performed using Cloud-Seg, a custom interactive software tool specifically developed for labeling large-scale 3D agricultural point clouds. Rather than labeling individual leaflets separately, all components of a single leaf were aggregated and annotated as one Compound Leaf-Node. The primary architectural axis was modeled as a sequence of Junction-Nodes connected by mainStem-Seg edges.

Graph-Based Structural Representation Graph representations were generated directly from the annotated point clouds using the companion Cloud-Graph pipeline. This method extracts a topology-preserving plant graph by identifying biologically meaningful nodes (e.g., the root node, axial junctions, and organ attachment points) and connecting them with edges representing the main stem segments, stalks, suckers, and compound leaves.

To ensure spatial consistency across the dataset, the input point cloud is normalized to a standardized, plant-centric coordinate frame. This is achieved by translating the root node to the origin (0,0,0) and aligning the primary vertical growth axis using Principal Component Analysis (PCA) [20]. Following spatial normalization, the points are grouped by their previously assigned semantic and instance identifiers to isolate discrete plant organs. For each structural segment, a discrete geometric representation of the organ's axis is generated by tracing a centerline path through a spatial neighborhood graph using shortest-path optimization algorithms [21].

The resulting plant graph successfully encodes both biological topology (node connectivity and branching order) and physical geometry (node spatial coordinates and edge centerlines). These graph networks are exported and stored in a lightweight, structured JSON format, serving as the crucial intermediate abstraction between the dense point clouds and the downstream extraction of phenotypic traits. Fig. 3 visualizes this structural encoding, demonstrating the precise spatial mapping achieved by the pipeline. The generated graph topology is overlaid onto the semi-transparent raw point cloud, with distinct architectural components (internodes, stalks, suckers, and leaves) and computed angular traits (insertion and phyllotaxy arcs) color-coded for visual clarity.

Extraction of phenotypic traits from topological graphs

4.1

Plant phenotypic traits were computed directly from the generated Cloud-Graph representations by interpreting each edge as a biologically meaningful plant element (e.g., internode, stalk, sucker, or compound leaf). Quantitative measurements were mathematically derived from the geometric centerline paths of these edges and the spatial vectors between connected nodes.

Whole-plant traits, including plant height and canopy dimensions, were computed from global geometric summaries of the graph in the normalized coordinate frame. Stem internodes were identified by ordering stem edges along the vertical axis between successive junction nodes, and internode length and curvature were computed from node-to-node distances and centerline geometry.

Lateral organs, including stalks and suckers, were analyzed relative to their attachment junctions. Leaf insertion angles were computed with respect to the local stem direction, while phyllotactic angles were derived as azimuthal divergence angles between consecutive lateral organs around the stem axis. These angles were computed using vector projections to ensure consistency across samples and developmental stages.

Internode length was computed as the Euclidean distance between consecutive junction-node coordinates (Pi, Pi+1) along the main stem: L_i_ = ||Pi - Pi+1||. For curved segments, geodesic length was computed by summing distances along the extracted centerline sample points(C), where C_k_ represents the 3D coordinate of the k-th sample: $[eqn]$ . Leaf insertion angle (α) was computed from the angle between the local stem direction vector (υstem) and the stalk/leaf direction vector (υleaf). $[eqn]$ Phyllotactic (divergence) angle was computed as the angle between consecutive insertion vectors projected onto the marker-defined XY plane. Angular traits are inherently scale-invariant and therefore robust to minor global scaling offsets introduced during photogrammetric reconstruction [22].

All extracted traits were exported to structured comma-separated values (CSV) files for downstream statistical analysis and model development. Optional visualization utilities allow the extracted graph and derived measurements to be overlaid on the original point cloud for qualitative verification.

Graph extraction, visualization, and phenotypic trait computation were implemented in Python (v3.11.11) [23]. using open-source scientific libraries. Core dependencies include Open3D (v0.19.0) for point-cloud processing and visualization [24], NumPy (v2.3.3) [25]. and SciPy (v1.16.2) for numerical operations [26], and scikit-learn (v1.7.2) for neighbourhood graph construction and auxiliary geometric processing [21].

Dataset utility

4.2

A core advantage of the TomatoPGT pipeline is its completely non-destructive nature, which allows continuous temporal tracking of plant architecture. Because the study emphasizes non-destructive workflows, no destructive manual measurements (e.g., harvesting organs for physical caliper measurement) were performed to cross-validate the extracted traits. Consequently, the dataset is explicitly framed as a high-fidelity “geometric ground truth” rather than an absolute “biological ground truth.” All provided phenotypic traits are deterministically derived from the validated metric geometry of the digital twin.

As established via the internal geometric validation, ICP benchmarking confirmed sub-millimeter metric stability across all scans, proving the geometric reliability of this ground truth. To ensure structural fidelity within this geometric space, all annotations and topological graphs were visually verified in 3D to confirm stem continuity, physiological junction placement, and lateral organ orientation.

By pairing dense semantic point clouds with structured topological graphs, TomatoPGT serves as a high-fidelity reference dataset for evaluating automated 3D plant analysis algorithms. We propose the following standard metrics for future benchmarking against this geometric ground truth:

• •Semantic & Instance Segmentation: Intersection over Union (IoU).
• •Graph Skeletonization: Chamfer Distance and Node Precision/Recall.
• •Phenotypic Trait Estimation: Mean Absolute Error (MAE) and Root Mean Square Error (RMSE).

Ultimately, TomatoPGT provides a rigorously validated foundation to accelerate computational agricultural research, explicitly supporting advancements in 3D segmentation, automated non-destructive phenotyping, and digital twin development.

Limitations

The 3D point clouds in this research were generated using a camera system mounted on a rotational arm, integrated with a turntable-based imaging setup. This configuration enabled precise and sequential 360° image capture around each plant but differs from fixed multi-camera systems such as those employing fifteen static cameras [10]. While this approach significantly reduces equipment cost and setup complexity, it may introduce minor spatial artifacts when the plant moves between frames due to rotation or vibration during image capture.

As the reconstructions rely on Structure-from-Motion and multi-view photogrammetry, geometric completeness may be reduced in highly concave or self-occluded canopy regions. Dense foliage intersections or overlapping leaves can introduce localized noise or partial reconstruction gaps, particularly in upper canopy areas. Nevertheless, the reconstructed point clouds provide sufficient geometric fidelity for organ-level annotation, structural graph extraction, and trait computation.

Computationally, the current implementation of the Cloud-Graph software has been validated on dense point clouds of up to approximately 5 million points using standard workstation hardware. This limitation is memory-bound rather than algorithmic; processing substantially larger datasets or scaling to field-level point clouds will require deployment on high-performance, high-memory computing clusters. Additionally, this foundational data descriptor does not include the evaluation of a baseline automated deep-learning or segmentation model.

In this study, plants were selected at similar developmental stages to ensure consistency during 3D reconstruction. Although three cultivars were included to introduce variation in morphology and leaf arrangement, additional genetic and growth diversity is recommended in future datasets to minimize potential bias and to further validate algorithmic generalization.

While the primary contribution of this work lies in providing a comprehensive 3D dataset (TomatoPGT) designed to evaluate the complete phenotyping pipelines, addressing the aforementioned limitations in future iterations will be highly beneficial. Future improvements may include expanding the imaging configuration to capture higher vertical angles and integrating real-time depth sensors to further enhance canopy completeness.

Ethics Statement

The authors declare that the ethical requirements of Data in Brief are considered. No human subjects, animals or data from any social media platform was included in this dataset.

CRediT Author Statement

Prasad Nethala: Conceptualization; methodology; software; data Annotation; data curation; validation; visualization; writing – original draft; writing – review & editing. Dugan Um: conceptualization; methodology; supervision; validation; visualization; writing – original draft; writing – review & editing. Samantha L. McCoy: Data curation; validation. Seth Gibson: Data curation; validation. Mahendra Bhandari: Supervision; resources; writing – review & editing. Kiju Lee: Supervision; resources; writing – review & editing