Dataset Card for CropAndWeed

This is a FiftyOne dataset with 8,034 samples.

Installation

If you haven't already, install FiftyOne:

pip install -U fiftyone

Usage

import fiftyone as fo
import fiftyone.utils.huggingface as fouh

dataset = fouh.load_from_hub("Voxel51/CropAndWeed")

session = fo.launch_app(dataset)

Dataset Details

Dataset Description

The CropAndWeed dataset is a large-scale, high-variability image dataset for Precision Agriculture, targeting the fine-grained identification of crop and weed species to support automated, plant-specific weed intervention. It provides bounding boxes, stem positions, and semantic segmentation masks for about 112k plant instances across more than 8k high-resolution real-world images, along with rich per-image meta-annotations describing recording conditions.

Images come from two complementary sources: an Application Set of images captured at several hundred commercially used agricultural sites in Austria (realistic, uncontrolled field conditions), and an Experimental Set captured on dedicated experimental plots specifically cultivated to enrich the dataset with otherwise underrepresented, rare weed species.

• Curated by: Daniel Steininger, Andreas Trondl, Gerardus Croonen, Julia Simon, and Verena Widhalm, AIT Austrian Institute of Technology
• Funded by: Austrian Federal Ministry for Climate Action, Environment, Energy, Mobility, Innovation and Technology, and the Austrian Research Promotion Agency (FFG), under the "ICT of the Future" research project HARV.EST (FFG No. 892347)
• Shared by: AIT Austrian Institute of Technology (original release); this FiftyOne format conversion was prepared by Harpreet Sahota
• Language(s): N/A (imagery, no text)
• License: Custom non-commercial research license (see Dataset Sources and More Information)

Dataset Sources

• Repository: https://github.com/cropandweed/cropandweed-dataset
• Paper: The CropAndWeed Dataset: A Multi-Modal Learning Approach for Efficient Crop and Weed Manipulation (WACV 2023)

Uses

Direct Use

• Training and benchmarking object detection, fine-grained plant classification, semantic/ panoptic segmentation, and stem/anchor-point (keypoint) localization models for Precision Agriculture and automated crop/weed manipulation.
• Academic research, teaching, and personal experimentation into robust, high-variability agricultural computer vision, including studying how incorporating multiple weed species into training affects crop-detection accuracy.
• Analysis of model robustness to environmental conditions using the per-image moisture, soil, lighting, and separability meta-annotations.

Out-of-Scope Use

• Any commercial use, including licensing, selling, or otherwise using the data (or the raw data itself in redistributed form) for commercial gain — explicitly disallowed by the dataset license.
• Redistributing the dataset or modified versions of it. Derivative works that are abstract representations not allowing recovery of the original data (e.g. trained model weights) are permitted; redistributing the images/annotations themselves is not.
• Use as a certified or standalone agronomic/regulatory plant-identification tool without expert validation — annotations were produced for computer-vision benchmarking purposes.

Dataset Structure

This is a flat (non-grouped) image dataset with media_type="image" and 8,034 samples, one sample per annotated image (1920x1088 RGB JPEG). It was parsed from the official raw annotation release (bboxes/, labelIds/CropAndWeed/, params/), preserving every modality shipped by the dataset authors.

Field	FiftyOne type	Description
filepath	StringField	Absolute path to the source RGB image
metadata	ImageMetadata	Standard image metadata (width, height, size_bytes, mime_type, num_channels)
recording_set	StringField	"Application Set" or "Experimental Set", parsed from the ave-/vwg- filename prefix
session	StringField	4-digit recording-session identifier (verbatim from filename)
image_id	StringField	4-digit image identifier within the session (verbatim from filename)
ground_truth	Detections	One Detection per annotated plant instance (2D bounding box), label is one of 100 fine-grained species/growth-stage classes (or the Vegetation fallback class)
stem_points	Keypoints	One Keypoint per annotated instance marking its stem (root) position
segmentation	Segmentation	Per-pixel semantic mask referencing a single-channel PNG (mask_path), same 100 classes as ground_truth
moisture / moisture_id	StringField / IntField	Soil moisture at capture time: dry, medium, or wet (raw id 0-2)
soil / soil_id	StringField / IntField	Soil granularity: fine, medium, or coarse (raw id 0-2)
lighting / lighting_id	StringField / IntField	Lighting condition: sunny or diffuse (raw id 0-1)
separability / separability_id	StringField / IntField	Subjective difficulty of separating plants from soil/each other: easy, medium, or hard (raw id 0-2)

Label types and why:

• ground_truth (Detections) — each row of the source bboxes/*.csv files (Left, Top, Right, Bottom, LabelID, StemX, StemY) is converted to a Detection with a relative [x, y, w, h] bounding box (required by FiftyOne). Each Detection also carries label_id (the raw numeric class id, 0-99 or 255) and stem_x/stem_y (the same stem position, kept as absolute pixel coordinates for reference).
• stem_points (Keypoints) — the dataset authors treat stem/anchor-point localization as a first-class learning task in its own right (Section 5.3 of the paper), so the stem position accompanying every bounding-box row is also materialized as a normalized-coordinate Keypoint, letting it be visualized and queried independently of ground_truth.
• segmentation (Segmentation) — the labelIds/CropAndWeed/*.png semantic masks are referenced via mask_path (not embedded as in-database arrays) to avoid duplicating ~8k full-resolution masks in the database. dataset.mask_targets["segmentation"] maps all 101 raw pixel values (0-99 plus 255) to their class names.
• Primitive fields for params — the four categorical per-image parameters are stored both as a human-readable string (moisture, soil, lighting, separability) and the raw numeric id shipped in params/*.csv (moisture_id, etc.), since the string versions are more convenient for filtering/grouping in the FiftyOne App.

dataset.classes — ground_truth and stem_points are both registered with the same 100-class list (99 fine-grained species/growth-stage labels plus the Vegetation fallback).

dataset.info contains the source repository URL, paper citation, and a short license summary.

Parsing decisions:

• Only the base, 100-class CropAndWeed label specification was imported. The official repository also defines simplified variants (Fine24, CropsOrWeed9, CropOrWeed2, Maize1/Maize2, SugarBeet1/SugarBeet2, etc.), which are pure label-ID remappings of these same raw annotations with no additional information — equivalent groupings can be reconstructed from ground_truth.detections.label_id using FiftyOne label-mapping views if needed, without re-importing.
• 234 of the 8,034 images have no annotated plant instances (their source bounding-box CSVs are empty); these samples have no ground_truth/stem_points but still have a segmentation mask (all background/Soil). All 8,034 images have a segmentation mask.
• label_id 255 (Vegetation) is a fallback class for tiny (<16² px), excluded, or visually ambiguous instances, and appears in both ground_truth and segmentation.
• No official train/val/test split is distributed with the raw annotation release (the paper's 70:15:15 split was generated for the authors' internal benchmarking only and is not shipped in the public download), so no split/tag field is set on this dataset.

Dataset Creation

Curation Rationale

Precision Agriculture approaches to automated weed intervention require large amounts of highly variable, annotated training data, which the authors found lacking in prior work (most existing datasets are small, laboratory-controlled, provide only one annotation modality, or cover far fewer species). CropAndWeed was curated to close this gap: a large-scale, high-variability dataset spanning 74 crop and weed species (16 crop, 58 weed), with multi-modal annotations (bounding boxes, semantic masks, and stem positions) rarely provided together elsewhere, and an emphasis on realistic field conditions plus specially cultivated plots for rare species.

Source Data

Data Collection and Processing

Images were captured over a period of four years (March-July) using a semi-professional full-frame SLR camera with a constant 50mm focal length, in auto-exposure mode, from an approximate top-down perspective at a height of about 1.1 meters. Each capturing session corresponds to a unique agricultural site or experimental plot at a specific point in time, yielding an average of about 20 images per session at intervals of at least three meters to avoid redundancy.

	Sessions (recorded / annotated)	Images (recorded / annotated)
Experimental Set	1,363 / 665	22,597 / 4,990
Application Set	738 / 264	21,217 / 3,044
Total	2,101 / 929	43,814 / 8,034

The 8,034 annotated images contain 111,953 annotated plant instances in total. Each recording session is also enriched with the moisture, soil, lighting, and separability meta-annotations included in this dataset, plus timestamps, GPS coordinates, and camera parameters recorded by the authors (not part of this FiftyOne conversion).

Who are the source data producers?

The AIT Austrian Institute of Technology research team (Daniel Steininger, Andreas Trondl, Gerardus Croonen, Julia Simon, Verena Widhalm), in cooperation with the experimental farm Groß-Enzersdorf, produced and recorded the source imagery.

Annotations

Annotation process

To increase annotation throughput, each image was first automatically pre-segmented into soil vs. vegetation pixels (initially via Otsu color thresholding, later via a CNN trained on already-annotated data). Annotators refined these masks and assigned the target species class. Bounding boxes were then auto-generated from the refined mask and manually corrected, and each plant's stem position was manually annotated. Image batches with high ambiguity or a high number of plant instances were reviewed by voting among multiple annotators as an additional quality-control step.

Who are the annotators?

The AIT annotation team, credited in the paper as Gulnar Bakytzhan, Vanessa Klugsberger, and Marlene Glawischnig, with additional support from Caroline Huber, Reinhard Neugschwandtner, and Helmut Wagentristl at the experimental farm Groß-Enzersdorf.

Personal and Sensitive Information

None. The dataset consists exclusively of top-down agricultural field imagery (soil, crops, and weeds); it contains no people, faces, or other personal or sensitive information.

Citation

BibTeX:

@InProceedings{Steininger_2023_WACV,
    author    = {Steininger, Daniel and Trondl, Andreas and Croonen, Gerardus and Simon, Julia and Widhalm, Verena},
    title     = {The CropAndWeed Dataset: A Multi-Modal Learning Approach for Efficient Crop and Weed Manipulation},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
    month     = {January},
    year      = {2023},
    pages     = {3729-3738}
}

APA:

Steininger, D., Trondl, A., Croonen, G., Simon, J., & Widhalm, V. (2023). The CropAndWeed dataset: A multi-modal learning approach for efficient crop and weed manipulation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (pp. 3729-3738).

More Information

The dataset is released by AIT Austrian Institute of Technology to academic and non-academic entities for non-commercial purposes only (academic research, teaching, scientific publications, or personal experimentation). Redistribution of the dataset or modified versions of it is not permitted; derivative works that are abstract representations not allowing recovery of the original data (e.g. trained models) are permitted. See the official license for the full terms.

The official repository additionally provides scripts to remap the raw 100-class annotations into 20 predefined simplified dataset variants (e.g. Fine24, CropsOrWeed9, CropOrWeed2) for applications that don't require the full fine-grained species/growth-stage taxonomy.

Dataset Card Authors

Harpreet Sahota (FiftyOne conversion and dataset card)