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/vineyard-pruning")

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

Dataset Details

Dataset Description

This dataset contains 536 vineyard images collected in Tierra de Barros, in the southwestern province of Badajoz, Extremadura, Spain, on Vitis vinifera vines of the "pardina"/"parda" and "macabeo"/"viura"/"lardot" varieties, grown in hedgerows under the "Ribera del Guadiana" designation of origin. Images were captured during manual pruning, performed by qualified professionals, under natural lighting and from varied angles, using a Canon EOS 2000D DSLR and a Samsung SM-A715F smartphone. Each image is annotated with polygon masks for three vine parts -- trunk, unpruned shoot, and pruned shoot -- labeled with VGG Image Annotator (VIA). The dataset has previously been used to train Mask R-CNN models for automatic localization of vine parts (reported mAP50 of 60.08), in support of automated/robotic vineyard pruning systems.

Curated by: Elia Pacioni, Eugenio Abengozar-García, Miguel Macías Macías, Carlos J. García Orellana, Horacio M. González Velasco, and Antonio García Manso (Universidad de Extremadura)
Funded by: Ministerio de Ciencia, Innovación y Universidades, Spain (Grant TED2021-131242B-I00, "Sistema robotizado para la poda inteligente de viñedos")
Shared by: Original data shared by the curators via Mendeley Data; this FiftyOne-formatted version prepared and shared by harpreetsahota
Language(s): Not applicable (image dataset; documentation/labels are in English)
License: CC BY 4.0

Dataset Sources

Repository: Vineyard dataset for pruning, Mendeley Data, V2 (DOI: 10.17632/n8cs4ns97p.2)
Paper: Pacioni, E., Abengózar, E., Macías Macías, M., García Orellana, C.J., González Velasco, H.M., & García Manso, A. (2025). Vineyard dataset for automatic pruning based on main parts localization. Data in Brief, 59, 111335. https://doi.org/10.1016/j.dib.2025.111335

Uses

Direct Use

This dataset is suitable for training and evaluating object detection / instance segmentation models (e.g. Mask R-CNN, or other polygon/mask-based detectors) to localize vine trunks and shoots for automated pruning robotics, viticulture research, and precision agriculture applications. It can also support studies of pruning-stage classification (distinguishing already-pruned from not-yet-pruned canes) given the per-image num_shoots / num_pruned_shoots counts included in this FiftyOne version.

Out-of-Scope Use

The dataset is limited to two grape varieties ("pardina" and "macabeo") grown in a single region (Tierra de Barros, Badajoz, Spain) under hedgerow training, so models trained on it may not generalize to other varieties, trellising systems, or geographies without further validation. It contains no disease, pest, or fruit/cluster annotations, and is not suitable for tasks outside vine-part localization (e.g. yield estimation, disease detection).

Dataset Structure

This is a flat image dataset (media_type="image") with 536 samples and no train/val/test splits (the source distributes a single unsplit pool).

Field	FiftyOne type	Description
`filepath`	`StringField`	Absolute path to the source JPEG image
`metadata`	`ImageMetadata`	Auto-computed width, height, MIME type, and file size
`ground_truth`	`Polylines`	One closed, filled `Polyline` per annotated vine-part instance, `label` in `{"shoot", "pruned shoot", "trunk"}`
`num_shoots`	`IntField`	Count of `"shoot"` (unpruned bud-stub) polylines in the image (derived)
`num_pruned_shoots`	`IntField`	Count of `"pruned shoot"` polylines in the image (derived)
`num_trunks`	`IntField`	Count of `"trunk"` polylines in the image (derived)
`num_regions`	`IntField`	Total polyline count for the image (derived)
`camera_make`	`StringField`	EXIF camera manufacturer, when available (412/536 images; 124 have no EXIF)
`camera_model`	`StringField`	EXIF camera model, when available
`capture_date`	`StringField`	ISO-8601 EXIF `DateTimeOriginal`, when available
`resolution`	`StringField`	`"{width}x{height}"` convenience string (5 distinct resolutions, 800x600 to 6000x4000)

dataset.info stores the original source URL, paper citation, license, a short description, and the classes list (["shoot", "pruned shoot", "trunk"]).

Label type choice: the source annotations are polygon masks (VIA shape_attributes.name == "polygon"), so each region maps to a FiftyOne Polyline (closed=True, filled=True) rather than a bounding-box Detection, preserving the original vertex-level segmentation boundary.

Parsing decisions:

VIA stores absolute pixel coordinates (all_points_x/all_points_y); these were converted to FiftyOne's normalized [0, 1] points using each image's actual on-disk width/height (the VIA JSON's size field is the file's byte size, not pixel dimensions, and was not used for this).
The VIA class attribute (region_attributes.clase, a Spanish-named radio field) maps "1"→"shoot", "2"→"pruned shoot", "3"→"trunk".
Samples were linked to their VIA record via the filename field inside each record, not the outer VIA dict key (which is an internal VIA dedup key concatenating the filename with part of a timestamp).
6,968 total polygon regions across 536 images (5,329 shoot / 887 pruned shoot / 752 trunk) -- 100% of images have at least one annotation (min 2, avg 13, max 30 regions/image).

Dataset Creation

Curation Rationale

Manual pruning is a labor-intensive, seasonal viticulture task, and labor scarcity is a recognized challenge for the Spanish agricultural sector. This dataset was curated to support development of an automated/robotic vine-pruning system by providing labeled imagery of the plant parts (trunk, unpruned shoot, pruned shoot) that such a system must localize.

Source Data

Data Collection and Processing

Images were collected in vineyards in Tierra de Barros, Badajoz, Extremadura, Spain (approx. 38.693, -6.409), on 10-20 year old, irrigated, hedgerow-trained Vitis vinifera vines ("pardina" and "macabeo" varieties) under the "Ribera del Guadiana" designation of origin. Photos were taken during manual pruning under natural lighting from varied angles, using a Canon EOS 2000D (34mm lens, sRGB, f/1.8) and a Samsung SM-A715F smartphone, producing five distinct resolutions from 800x600 to 6000x4000. The source paper states images were collected across 2021-2023; EXIF DateTimeOriginal values recovered during FiftyOne ingestion show only 2022 (Samsung batch, 261 images) and 2024 (Canon batch, 151 images), with 124 images (the lower-resolution buckets) carrying no EXIF data at all -- noted here for transparency, as it was not resolved from the source materials.

Who are the source data producers?

Images were captured by the curating research team (Universidad de Extremadura, Instituto de Computación Científica Avanzada) during pruning work carried out by qualified vineyard pruning professionals.

Annotations

Annotation process

Each image was manually labeled with polygon masks using VGG Image Annotator (VIA). Three region classes were defined: trunk (the vine trunk, present in nearly every image), unpruned shoot (only the first bud-bearing segment of an as-yet-unpruned cane is labeled, not its full length), and pruned shoot (shoots that have already been cut). Per the source paper, label accuracy was validated with ad-hoc scripts in both VIA and COCO format, though only the VIA project file is distributed with the dataset.

Who are the annotators?

The dataset curators (Universidad de Extremadura team listed under Curated By).

Personal and Sensitive Information

None identified. Images depict vineyard plant material (trunks and shoots); no people, faces, or other personally identifying information are annotated or apparent in the sampled imagery.

Citation

BibTeX:

@article{pacioni2025vineyard,
  title   = {Vineyard dataset for automatic pruning based on main parts localization},
  author  = {Pacioni, Elia and Abeng{\'o}zar, Eugenio and Mac{\'i}as Mac{\'i}as, Miguel and Garc{\'i}a Orellana, Carlos J. and Gonz{\'a}lez Velasco, Horacio M. and Garc{\'i}a Manso, Antonio},
  journal = {Data in Brief},
  volume  = {59},
  pages   = {111335},
  year    = {2025},
  doi     = {10.1016/j.dib.2025.111335}
}

@misc{pacioni2024vineyarddata,
  title     = {Vineyard dataset for pruning},
  author    = {Pacioni, Elia and Abengozar-Garc{\'i}a, Eugenio and Mac{\'i}as Mac{\'i}as, Miguel and Garc{\'i}a Orellana, Carlos J. and Gonz{\'a}lez Velasco, Horacio M. and Garc{\'i}a Manso, Antonio},
  year      = {2024},
  publisher = {Mendeley Data},
  version   = {V2},
  doi       = {10.17632/n8cs4ns97p.2}
}

APA:

Pacioni, E., Abengózar, E., Macías Macías, M., García Orellana, C. J., González Velasco, H. M., & García Manso, A. (2025). Vineyard dataset for automatic pruning based on main parts localization. Data in Brief, 59, 111335. https://doi.org/10.1016/j.dib.2025.111335

More Information

This FiftyOne-formatted version was produced by parsing the source VIA (vidprune.json) project file directly, since VIA is not one of FiftyOne's built-in importer formats. All 536 images and 6,968 polygon regions were verified to match the counts reported in the source paper's Table 1 and Figure 2 during ingestion.

Dataset Card Authors

Harpreet Sahota

Downloads last month: -

Total file size:

3.21 GB