Voxel51/PhenoBench_Raw

Dataset Card for PhenoBench

This is a FiftyOne dataset with 2872 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/PhenoBench_Raw")

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

Dataset Details

Dataset Description

PhenoBench is a large dataset for the semantic interpretation of images of real agricultural fields. It provides dense, pixel-wise annotations of a sugar beet field recorded with a UAV, covering semantic segmentation of crops/weeds/soil, plant-level instance segmentation, and crop leaf-level instance segmentation, all captured across three growth stages during the 2020 growing season. The dataset also supports a novel "hierarchical panoptic segmentation" task that jointly reasons about plants and their constituent leaves. A held-out test set (including images from a second, unseen field recorded in 2021) is scored via a server-side benchmark, so its ground truth is not distributed.

This FiftyOne dataset was built directly from the official PhenoBench-v110.zip release, preserving the original train/val/test splits and all pixel-wise annotations.

• Curated by: Jan Weyler, Federico Magistri, Elias Marks, Yue Linn Chong, Matteo Sodano, Gianmarco Roggiolani, Nived Chebrolu, Cyrill Stachniss, and Jens Behley (Center for Robotics, University of Bonn; University of Oxford; Lamarr Institute for Machine Learning and Artificial Intelligence)
• Funded by: Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany's Excellence Strategy, EXC-2070 – 390732324 (PhenoRob)
• Shared by: Harpreet Sahota (FiftyOne conversion); original dataset by the PhenoBench authors (University of Bonn)
• Language(s): en (not applicable to image content; metadata/documentation is in English)
• License: CC BY-SA 4.0, per the official dataset page. Note: the README.MD bundled inside the downloaded PhenoBench-v110.zip (v1.1.0) instead states CC BY-NC-SA 4.0, citing the earlier 2023 arXiv preprint. This card follows the current, live license terms published on the official website, which cites the 2024 T-PAMI publication. If you need certainty for commercial use, confirm directly with the authors (contact below).

Dataset Sources

• Repository: github.com/PRBonn/phenobench (devkit), github.com/PRBonn/phenobench-baselines (baselines)
• Paper: PhenoBench — A Large Dataset and Benchmarks for Semantic Image Interpretation in the Agricultural Domain, IEEE T-PAMI, 46(12), 9583–9594, 2024
• Demo: phenobench.org

Uses

Direct Use

• Semantic segmentation of crop / weed / soil pixels
• Plant-level instance segmentation and detection (crops and weeds)
• Crop leaf-level instance segmentation and detection
• Hierarchical panoptic segmentation (joint plant + leaf granularity)
• Studying occlusion/visibility effects using the per-instance visibility scores
• Growth-stage / temporal analysis using the acquisition_date field (three recording dates roughly a week apart)
• Benchmarking against the official hidden test-set leaderboards hosted via phenobench.org/benchmarks.html

Out-of-Scope Use

• The dataset covers a single ~1,300 m² sugar beet field with two crop varieties (BTS 440, Celesta KWS) and six regional weed species observed in Germany; models trained here may not generalize to other crops, weed floras, or geographies without further validation.
• The test split in this FiftyOne dataset intentionally has no ground_truth_* labels (they are withheld for the official server-side evaluation), so it cannot be used for local supervised training or offline validation — only for generating predictions to submit to the benchmark server.
• Not intended for any use requiring identification of people; the dataset contains no personal or sensitive information (see below).

Dataset Structure

This is a flat (non-grouped) FiftyOne image dataset with media_type="image" and 2,872 samples, corresponding exactly to the official train (1,407), val (772), and test (693) splits of PhenoBench-v110.zip.

Field	FiftyOne type	Description
filepath	StringField	Path to the RGB image (1024×1024, verbatim from source)
tags	ListField(StringField)	Set to the split name (train, val, or test)
split	StringField	Dataset split: train, val, or test (verbatim from source directory structure)
acquisition_date	StringField	ISO date the UAV image was recorded — 2020-05-15, 2020-05-26, or 2020-06-05; None for the test split, whose filenames are anonymized (phenoBench_XXXXX.png)
metadata	ImageMetadata	Standard FiftyOne image metadata (width, height, num_channels, size_bytes, mime_type), computed via dataset.compute_metadata()
ground_truth_semantic	Segmentation	Pixel-wise semantic mask, referencing the original 16-bit semantics/*.png directly via mask_path. Classes: background (0), crop (1), weed (2), partial_crop (3), partial_weed (4). None for test
ground_truth_plants	Detections	One Detection per plant instance (crop or weed), derived from plant_instances/, semantics/, and plant_visibility/. None for test
ground_truth_leaves	Detections	One Detection per crop leaf instance, derived from leaf_instances/ and leaf_visibility/. None for test

Each Detection in ground_truth_plants and ground_truth_leaves additionally carries:

Attribute	Type	Description
label	string	crop, weed, partial_crop, or partial_weed for plants; always "leaf" for leaves
mask	boolean array	Per-instance segmentation mask, cropped to the instance's bounding box
visibility	float, 0–1	Fraction of the instance visible in-frame, derived from plant_visibility/leaf_visibility (0–255 in the source, normalized here). Instances with a mostly-visible plant are labeled crop/weed; instances with <50% visible pixels are labeled partial_crop/partial_weed
source_instance_id	int	The original (arbitrary, non-sequential) instance id from plant_instances/leaf_instances
source_plant_instance_id	int	(leaves only) The source_instance_id of the parent plant instance a leaf belongs to

dataset.info["field_location"] stores the single physical location all images were recorded at (not a per-sample field, since every image comes from the same field):

{
    "description": "All images were captured via UAV from a single ~1300 m^2 sugar beet field",
    "site": "Campus Klein-Altendorf farm, University of Bonn, between Meckenheim and Rheinbach, Germany",
    "latitude": 50.6252,
    "longitude": 6.9887,
    "coordinates_dms": "50°37'.51\"N, 6°59'.32\"E",
}

Parsing decisions

• Instance → Detection conversion. By directly inspecting the pixel data, we verified that every plant_instances id maps to exactly one semantic class and one constant plant_visibility value across all of its pixels, and every leaf_instances id belongs to exactly one parent plant instance (leaves are only annotated for crops, never weeds). This let us convert the raw instance-id arrays into one Detection per instance (using scipy.ndimage.find_objects for efficient bounding-box/mask extraction) rather than keeping them as raw index-array Segmentation fields, which makes individual plants/leaves filterable, sortable by visibility, and clickable in the FiftyOne App.
• Leaf labels. All leaf detections use a flat "leaf" label rather than inheriting the parent plant's crop/partial_crop class, since leaves are exclusively annotated for crop plants.
• Semantic masks kept as file references. ground_truth_semantic uses mask_path pointing at the original semantics/*.png files rather than duplicating the mask data, since the source PNGs are already in a directly-usable single-channel format.
• Test split. The test split's ground_truth_* fields are None by design — labels are withheld by the dataset authors for their server-side benchmark evaluation.
• Validation performed. Sample and instance counts were independently cross-checked against a raw scan of the source PNGs (30,424 plant instances, 106,767 leaf instances — exact match) and against the published dataset statistics in the paper (Table 2: train+val crop/weed/leaf counts match exactly once partial-visibility subclasses are combined with their full-visibility counterparts).

Dataset Creation

Curation Rationale

Dense, pixel-accurate agricultural datasets with reproducible, hidden-test-set benchmarks were largely missing prior to PhenoBench (see Table 1 of the paper for a comparison against CWFID, CVPPP, WeedMap, GrowliFlowers, CropAndWeed, and others). The authors set out to provide a large-scale dataset that simultaneously supports semantic segmentation, plant-level instance/panoptic segmentation, and — uniquely — crop leaf-level instance segmentation, enabling a new "hierarchical panoptic segmentation" task that reflects how plant scientists and breeders actually assess plant growth stage (e.g., leaf count per plant).

Source Data

Data Collection and Processing

Images were recorded with a DJI M600 UAV equipped with a PhaseOne iXM-100 camera (80 mm RSM prime lens on a gimbal), flying at ~21 m altitude for a ground sampling distance (GSD) of ~1 mm/px, with 75% forward and 50% side image overlap; each image is geo-referenced via the onboard GNSS. Three flights were performed roughly a week apart in 2020 (May 15, May 26, June 5) over an ~1,300 m² sugar beet field at the Campus Klein-Altendorf farm (University of Bonn), between Meckenheim and Rheinbach, Germany (50°37'.51"N, 6°59'.32"E). Eight crop rows covered by the mission were spatially split into 4 rows for training, 2 for validation, and 2 for testing, ensuring no train/test spatial leakage. Four additional 2021 flights over a second, unseen field were folded only into the test split, to evaluate generalization to novel fields. The field itself contains two sugar beet varieties (BTS 440, Celesta KWS) and six naturally occurring weed species (Chenopodium album, Polygonum aviculare, Thlaspi arvense, Persicaria lapathifolia, Bilderdykia convolvulus, Polygonum hydropiper), recorded under varying herbicide regimes (fully, partially, and non-herbicided) and lighting conditions (sunny and overcast).

Full-resolution UAV captures (11,664×8,750 px) were tiled into overlapping 2,000×2,000 px patches across four iterations to guarantee every plant was completely visible in at least one patch, for annotation purposes. The final released images/annotations were then re-extracted from the full-resolution captures as 1,024×1,024 px tiles with 50% overlap, sized so that even later growth-stage plants (with more overlap between neighboring plants) remain mostly visible within a single tile.

Who are the source data producers?

The source data is real agricultural field imagery from an experimental farm operated by the University of Bonn (Campus Klein-Altendorf), not crowd-sourced or synthetic. No human subjects are depicted.

Annotations

Annotation process

Annotation used a custom hierarchical labeling tool built by the authors. Plant instances (crops and weeds) were labeled first by 9 annotators (approx. 800 h total), with each of the four tiling iterations validated and corrected before annotations were transferred back onto the full-resolution image and the next iteration begun. Leaf instances were labeled next by 5 annotators (approx. 1–2 h per patch), constrained by the tool to lie inside their parent plant's mask, which is what allows each leaf to be reliably associated with its parent plant instance. A further round of correction/validation was performed by 4 additional annotators. In total, 14 annotators invested approx. 1,400 h of direct annotation plus approx. 600 h of validation/refinement (approx. 2,000 h overall).

Crop plant instances also carry a temporally-consistent id across the three 2020 recording dates: camera poses were estimated via RTK-GNSS-initialized bundle adjustment, and crop centroids were associated across dates using Hungarian matching (15 cm distance threshold), refined with a small number of manually-associated anchor plants per date pair. This associated 583 unique crop plants over the season (496 observed on all three dates). Note: this cross-date instance linkage is part of the original data/paper but is not separately exposed as a FiftyOne field in this conversion — only per-image instance ids are retained (see source_instance_id).

Who are the annotators?

Students at the University of Bonn, supervised by the paper's authors ("We thank all students annotating the data.").

Personal and Sensitive Information

None. The dataset consists exclusively of UAV imagery of soil and plants in an agricultural field; it contains no people, faces, personal identifiers, or other personal/sensitive information.

Citation

BibTeX:

@article{weyler2024pami,
  author = {Jan Weyler and Federico Magistri and Elias Marks and Yue Linn Chong and Matteo Sodano
    and Gianmarco Roggiolani and Nived Chebrolu and Cyrill Stachniss and Jens Behley},
  title = {{PhenoBench --- A Large Dataset and Benchmarks for Semantic Image Interpretation
    in the Agricultural Domain}},
  journal = {IEEE Trans. on Pattern Analysis and Machine Intelligence (T-PAMI)},
  volume = {46},
  number = {12},
  pages = {9583--9594},
  year = {2024}
}

APA:

Weyler, J., Magistri, F., Marks, E., Chong, Y. L., Sodano, M., Roggiolani, G., Chebrolu, N., Stachniss, C., & Behley, J. (2024). PhenoBench — A Large Dataset and Benchmarks for Semantic Image Interpretation in the Agricultural Domain. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12), 9583–9594.

More Information

• The official release also provides a large amount of unlabeled UAV imagery (patches and rotation-augmented patches, ~129,000 images across 7 additional recording dates spanning April–July 2020) for self-supervised pre-training; these are distributed separately on phenobench.org and are not included in this FiftyOne dataset, which only covers the labeled PhenoBench-v110.zip release.
• A Python devkit (pip install phenobench) with a PyTorch dataloader, visualization helpers, and the official evaluation/validator scripts (phenobench-eval, phenobench-validator) is available at github.com/PRBonn/phenobench.
• For questions about commercial use or licensing, the original authors can be reached at photogrammetry@uni-bonn.de.

Dataset Card Authors

Harpreet Sahota (FiftyOne conversion and dataset card)