Voxel51/tomato-map

Dataset Card for TomatoMAP

This is a FiftyOne dataset with 68,069 samples.

Installation

If you haven't already, install FiftyOne:

pip install -U fiftyone

Usage

import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub

# Load the dataset
# Note: other available arguments include 'max_samples', etc
dataset = load_from_hub("Voxel51/tomato-map")

# Launch the App
session = fo.launch_app(dataset)

Dataset Details

Dataset Description

TomatoMAP is a multi-angle, multi-pose, multi-task imaging dataset of Solanum lycopersicum (tomato) collected with a purpose-built IoT imaging station at the Julius Kühn-Institute (JKI), Quedlinburg, Germany. It unifies three originally-separate TomatoMAP subsets -- TomatoMAP-Cls (BBCH growth-stage classification), TomatoMAP-Det (region-of-interest object detection), and TomatoMAP-Seg (fruit/flower developmental-stage instance segmentation) -- into a single grouped FiftyOne dataset, tagged by source.

101 tomato plants (all the "Money Maker" accession, non-transgenic) were imaged over a 163-day period by 4 fixed Raspberry Pi cameras (mounted at 45/90/135/180 degree elevation angles) while a turntable rotated the plant through 12 poses (30 degree increments), giving 48 synchronized images per plant per acquisition session -- 64,464 images in total, each with a 7-class ROI bounding-box annotation and a session-level BBCH growth-stage label (50 distinct codes). A separate, later imaging pass with a handheld macro camera produced 3,605 high-resolution close-up photographs of individual flower buds and fruit clusters, 727 of which have ISAT-tool instance segmentation masks across 10 fruit/flower developmental-stage classes.

• Curated by: Yujie Zhang, Sabine Struckmeyer, Andreas Kolb, Sven Reichardt (Julius Kühn-Institute & University of Siegen); parsed into FiftyOne format by Harpreet Sahota.
• Funded by: German Federal Ministry of Agriculture, Food and Regional Identity; compute powered by the de.NBI Cloud / ELIXIR-DE.
• Shared by: Harpreet Sahota (FiftyOne format on Hugging Face); original data deposited by the authors in e!DAL (IPK Gatersleben).
• Language(s): en (metadata, class names, and the BBCH index descriptions are in English)
• License: CC BY 4.0

Dataset Sources

• Repository: e!DAL / IPK Gatersleben (raw data); github.com/0YJ/TomatoMAP (original TomatoMAP-Cls/Det builder + model code)
• Paper: Zhang et al., "Tomato Multi-Angle Multi-Pose Dataset for Fine-Grained Phenotyping," Scientific Data (2026) 13:309
• Demo: 0yj.github.io/tomato_map

Uses

Direct Use

• Training/benchmarking fine-grained BBCH phenological growth-stage classifiers.
• Training/benchmarking object detectors for plant ROIs (leaf, whole plant, panicle, flower clusters, fruit clusters, axillary shoot, shoot), including studying class imbalance and overlapping-box handling.
• Training/benchmarking instance/semantic segmentation of fruit ripening stages (nascent -> mini -> unripe -> semi ripe -> fully ripe) and flower/bud size stages (2mm -> 4mm -> 6mm -> 8mm -> 12mm).
• Multi-view / sparse-view 3D reconstruction research: the 4 rig cameras have known intrinsics (data/calibration/pi{1-4}.npz) and nominal elevation/turntable angles, though no extrinsic rig calibration is provided.
• Longitudinal growth-stage analysis using capture_datetime and plant_id across the 163-day acquisition window.

Out-of-Scope Use

• Disease/pathogen detection: all plants are healthy specimens of a single cultivar ("Money Maker"), non-transgenic -- there is no disease-state variation to learn from (unlike e.g. PlantVillage or Tomato-Village).
• Cross-referencing individual plants or sessions between the det/cls samples and the seg samples: they come from different camera systems, different (only lightly overlapping) capture windows, and carry no shared plant/session identifier -- see dataset.info["seg_source"] for the evidence. Do not assume a plant_id-level relationship between the two subsets.
• Genotype/accession comparison studies: every plant in the det/cls subset is the same accession, grown in the same cabin, so there is no genetic or environmental variation encoded in the metadata.

Dataset Structure

This is a single grouped FiftyOne dataset (media_type="group", group field group) with 68,069 samples across 19,721 groups, unifying two source subsets that are tagged but not otherwise sample-joinable (see Out-of-Scope Use):

• det (64,464 samples, tag "det") -- the TomatoMAP-Cls/Det rig images. Grouped by image_set_id (one turntable pose held still, shot simultaneously by the 4 fixed cameras), with slices "pos_1".."pos_4" corresponding to the 4 camera elevations. 16,116 groups x 4 slices = 64,464 samples.
• seg (3,605 samples, tag "seg") -- the TomatoMAP-Seg macrophotographs. Each image is its own group (no natural grouping partner), populating a single slice "macro". Further tagged "labeled" (727 samples with ISAT annotations) or "unlabeled" (2,878 samples, bonus un-annotated photos).

Fields

Field	FiftyOne type	Subset	Description
filepath	StringField	both	Path to the image file
tags	ListField(StringField)	both	"det"/"seg" (source subset) plus "labeled"/"unlabeled" (seg only)
group	Group	both	FiftyOne group field; .name is the slice (pos_1..pos_4 or macro)
metadata	ImageMetadata	both	Standard FiftyOne image metadata (width/height/size/mime type)
ground_truth	Detections	both	Det: 7-class YOLO ROI boxes. Seg: instance masks over 10 fruit/flower-stage classes (each Detection also carries isat_group (int instance id), area_px, iscrowd). One shared field name since the two class vocabularies never overlap.
classification	Classification	det	BBCH growth-stage label, e.g. "bbch_70" -- one label per (plant_id, capture date) session, inherited by all 48 images from that session
bbch_stage	IntField	det	Raw BBCH code (13-89; 50 distinct codes appear in the data)
bbch_description	StringField	det	Human-readable description of the BBCH code
image_set_id	IntField	det	Groups the 4 simultaneous camera exposures of one turntable pose (verbatim from metadata/raw_pheno.csv / parsed from filename)
plant_id	IntField	det	Plant identifier, 1-101
piid	IntField	det	Rig camera/position id, 1-4
pose_id	IntField	det	Turntable pose id, 1-12
pose_degrees	IntField	det	Turntable rotation in degrees (0-330, 30 degree steps)
camera_label	StringField	det	Human-readable camera description, e.g. "bottom (45 deg)"
camera_angle_deg	IntField	det	Camera elevation angle code (verbatim from metadata/camera.csv)
camera_position_desc	StringField	det	Camera position description (verbatim from metadata/camera.csv)
accession	StringField	det	Plant accession (constant: "Money Maker")
transgene	BooleanField	det	Transgenic status (constant: False)
cabin	StringField	det	Greenhouse cabin id (constant: "H01504")
capture_datetime	DateTimeField	both	Capture timestamp -- det: parsed from the filename's 14-digit timestamp; seg: parsed from the image's EXIF DateTimeOriginal
f_number, exposure_time, iso, focal_length, camera_model	FloatField/StringField/IntField	seg	EXIF capture settings for the macro camera (verbatim from metadata/TomatoMAP-Seg_meta.csv)

dataset.classes["ground_truth"] holds all 17 classes (7 det ROIs + 10 seg stages); dataset.classes["classification"] holds the 50 bbch_* labels present in the data.

dataset.info

Dataset-level metadata (not attached to individual samples): paper/dataset DOIs, code and project-homepage links, license, a description of each subset's imaging setup and capture window, a pointer to the (sample-unattached) camera calibration files, and known data-quality notes -- including that 1 of the 64,464 det images has no YOLO label file, and that the source paper's Table 3 lists 91,120 instances for BBCH 80-89 while the exact count in the released data is 9,120 (the other two coarse buckets, 10,560 and 29,328, match the paper exactly, and 10,560 + 29,328 + 9,120 sums with the unclassified/other-stage images to the full 64,464 -- this looks like a typo in the publication, not a data issue).

Saved views

View	Samples	Description
det	64,464	The det/cls subset
seg	3,605	The seg subset
seg-labeled	727	Seg images with ISAT annotations
seg-unlabeled	2,878	Seg images with no annotation on file
bbch-vegetative-flowering	10,560	Det images with BBCH stage 60-69
bbch-flowering	29,328	Det images with BBCH stage 70-79
bbch-fruit-development	9,120	Det images with BBCH stage 80-89
det-missing-labels	1	QA view: det images with zero ROI detections
camera-pos_1 .. camera-pos_4	16,116 each	One fixed rig camera elevation each

Indexes

plant_id, image_set_id, piid, pose_id, bbch_stage, classification.label, ground_truth.detections.label, capture_datetime -- chosen for the fields most useful to filter/sort/group by, on top of FiftyOne's default indexes (id, filepath, group.id, group.name, tags, etc.). Constant fields (accession, transgene, cabin) are intentionally not indexed.

Parsing decisions

• BBCH is a session-level label, not per-image. metadata/BBCH_classification.xlsx gives one BBCH code per (plant_id, calendar date); every image captured in that plant's 12-pose x 4-camera session on that date inherits the same classification value.
• metadata/raw_pheno.csv only covers ~12% of images (7,776 of 64,464, 61 of 101 plants) -- it was used as an authoritative cross-check where available, but filename parsing (pi{cam}_{seq}_{plant}_{pose}_{timestamp}.jpg) is the primary, and only complete, source of the plant_id/piid/pose_id/image_set_id/capture_datetime fields.
• Seg masks are rasterized from ISAT polygon annotations (TomatoMAP-Seg/labels/*.json) into per-detection boolean masks cropped to each object's bounding box.
• Seg and det/cls are treated as unrelated pools on purpose -- see Out-of-Scope Use above.

Dataset Creation

Curation Rationale

Observer bias and inconsistency in manual plant phenotyping limit the accuracy and reproducibility of fine-grained trait analysis. TomatoMAP was built to provide a large, standardized, multi-view imaging dataset -- captured under a fixed IoT-based acquisition protocol -- to train and validate real-time, accuracy/efficiency-balanced computer vision models (MobileNetv3, YOLOv11, Mask R-CNN in the original paper) as a substitute for manual phenotyping, and to quantify how closely such models agree with human experts (via Cohen's Kappa and inter-rater agreement analysis).

Source Data

Data Collection and Processing

Det/Cls images were captured by a custom data-acquisition station: 4 OV5647 5MP color CMOS cameras (three with 90 deg lenses, one with a 170 deg fisheye lens) mounted at 45/90/135/180 degree vertical inclination, aimed at a turntable that rotated a potted plant through 12 poses at 30 degree increments, synchronized across all 4 cameras at each rotational step. 101 plants were imaged this way over a 163-day period (2023-08-16 to 2024-01-26 in the released data, irregular intervals of 1-13 days), yielding 64,464 images at 1080x1440 resolution. Camera intrinsics/distortion were calibrated with a planar chessboard pattern.

Seg macrophotographs were captured separately with a Panasonic Lumix DMC-FZ1000 at 3648x5472 resolution, over a different (later, mostly non-overlapping) date range -- the ISAT annotation files' embedded folder paths (MPTSTD_dataset_boost/task{1,2,3}) indicate this was its own imaging task, not the same rig/plant cohort as the det/cls subset.

Who are the source data producers?

Researchers at the Julius Kühn-Institute (Institute for Breeding Research on Horticultural Crops, Quedlinburg) and the Computer Graphics Group, University of Siegen, using an automated greenhouse imaging system of their own design.

Annotations

Annotation process

• Det (ROI boxes): a progressive, AI-assisted labeling workflow. An initial 1,780 images were manually labeled in Label Studio to train a first assistive model; that model pre-labeled 2,504 more images, which were expert-reviewed/corrected and merged in to train a second assistive model; a third round applied the second model to a 6,000-image pool with further expert validation; all annotation files were then cross-checked by five experts. Bounding boxes were drawn to tightly enclose visible extent (plus morphologically plausible occluded extent for partially visible objects); intra-class overlap over 70% was disallowed, cross-class overlap (e.g. panicle containing flower/fruit clusters) was permitted.
• Cls (BBCH stage): assigned per (plant, date) session according to the standardized BBCH developmental scale for vegetables/fruiting crops.
• Seg (instance masks): annotated with the Interactive Semi-Automatic Annotation Tool (ISAT), using Segment Anything Model 2 (SAM2) for proposal generation, followed by manual refinement, producing pixel-wise polygon masks with per-instance group ids.

Who are the annotators?

Domain experts at JKI, with annotations cross-validated among five annotators (self-inspection, senior-annotator spot review of >=10% of samples, and cross-annotator validation), plus AI-vs-human agreement analysis against 5 named domain experts in the source paper.

Personal and Sensitive Information

None. The dataset contains only plant imagery and greenhouse/imaging-hardware metadata; no personal or human-subject data is present.

Citation

BibTeX:

@article{zhang2026tomatomap,
  author  = {Zhang, Yujie and Struckmeyer, Sabine and Kolb, Andreas and Reichardt, Sven},
  title   = {Tomato Multi-Angle Multi-Pose Dataset for Fine-Grained Phenotyping},
  journal = {Scientific Data},
  year    = {2026},
  volume  = {13},
  pages   = {309},
  doi     = {10.1038/s41597-026-06926-9}
}

@misc{zhang2025tomatomapdata,
  author    = {Zhang, Yujie and Struckmeyer, Sabine and Kolb, Andreas and Reichardt, Sven},
  title     = {Tomato Multi-Angle Multi-Pose Dataset for Fine-Grained Phenotyping},
  year      = {2025},
  publisher = {e!DAL -- Leibniz Institute of Plant Genetics and Crop Plant Research (IPK)},
  doi       = {10.5447/ipk/2025/14}
}

APA:

Zhang, Y., Struckmeyer, S., Kolb, A., & Reichardt, S. (2026). Tomato Multi-Angle Multi-Pose Dataset for Fine-Grained Phenotyping. Scientific Data, 13, 309. https://doi.org/10.1038/s41597-026-06926-9

More Information

The original TomatoMAP-Cls/Det builder notebooks and model training/evaluation code are at github.com/0YJ/TomatoMAP. This FiftyOne version was built directly from the raw rig images + YOLO labels + ISAT segmentation JSONs (not from the authors' pre-built train/val/test split, since that split was generated with a random seed the authors' code controls, not distributed as a fixed file) -- see dataset.info and the saved views above for how to reconstruct comparable subsets.

Dataset Card Authors

Harpreet Sahota (FiftyOne format and this card)