Note

This is a Hugging Face dataset. For large datasets, ensure huggingface_hub>=1.1.3 to avoid rate limits. Learn more in the Hugging Face integration docs.

Hugging Face

Dataset Card for CottonWeedDet12#

image/png

CottonWeedDet12 is a 12-class weed object-detection dataset for cotton production systems in the southern U.S., consisting of 5,648 RGB field images with 9,370 bounding box annotations collected in Michigan State University MEFAS field trials during 2021-2022. It is the companion dataset for the YOLOWeeds benchmark of YOLO object detectors.

This is a FiftyOne dataset with 5648 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/CottonWeedDet12")

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

Dataset Details#

Dataset Description#

Weeds are among the major threats to cotton production; overreliance on herbicides has accelerated herbicide-resistance in weeds and raised concerns about environmental and food safety impacts. CottonWeedDet12 was built to support machine-vision, robotic-weeding research by providing a large, annotated, multi-class image dataset of the weeds most important to cotton production in the southern United States. 5,648 RGB photos were collected with smartphones and hand-held digital cameras under natural field lighting, at varied weed growth stages, across Michigan State University MEFAS field sites between June and September 2021 (a small batch of 34 images in the shipped download is dated January 2022 — see the Discrepancies note under More Information). Each image was manually annotated with bounding boxes for one or more of 12 weed classes using the VGG Image Annotator (VIA) v2.10, yielding 9,370 boxes total.

  • Curated by: Yuzhen Lu, Michigan State University (with co-authors Fengying Dang, Dong Chen, Zhaojian Li)

  • Funded by: [More Information Needed]

  • Shared by: Yuzhen Lu, Michigan State University (Zenodo upload)

  • Language(s): N/A (image dataset; no text content)

  • License: cc-by-nc-4.0

Dataset Sources#

  • Repository: https://zenodo.org/records/7535814 (original release); companion code: https://github.com/DongChen06/DCW

  • Paper: Dang, F., Chen, D., Lu, Y., Li, Z. (2023). YOLOWeeds: A novel benchmark of YOLO object detectors for multi-class weed detection in cotton production systems. Computers and Electronics in Agriculture, 205, 107655. https://doi.org/10.1016/j.compag.2023.107655

  • Demo: [More Information Needed]

Uses#

Direct Use#

Training and benchmarking object-detection models (e.g. YOLO family detectors) for multi-class weed identification and detection in cotton fields; research on machine-vision systems for automated/robotic weed control and integrated, sustainable weed management in cotton production.

Out-of-Scope Use#

Not validated for weed detection in crops, regions, or growth conditions outside southern U.S. cotton production systems without further evaluation. The cc-by-nc-4.0 license prohibits commercial use without separate permission from the rights holder.

Dataset Structure#

This is a flat, non-grouped image dataset (media_type = "image") with 5,648 samples and no train/val/test splits (none were shipped with the source data). Each sample is a single RGB photograph of one or more weeds, annotated with 2D bounding boxes.

Field

FiftyOne type

Description

filepath

StringField

Path to the source RGB image (.jpg, variable resolution)

ground_truth

Detections

12-class weed bounding boxes, one Detection per annotated weed instance

capture_date

StringField

Capture date (YYYYMMDD), parsed from the filename (verbatim from source naming convention)

device

StringField

Capture device (e.g. iPhoneSE, NIKOND3300, iPhone11Pro, CanonEOS4000D, YALAL00), parsed from the filename

annotator_initials

StringField

Photographer/annotator initials, parsed from the filename

Label type and why: the source annotations are 2D rectangular bounding boxes (one or more per image), so they map directly to FiftyOne’s Detections type under the ground_truth field. The 12 weed classes (with box counts in this release) are:

Class

Box count

Waterhemp

1,959

MorningGlory

1,341

Purslane

992

SpottedSpurge

952

Carpetweed

962

Ragweed

896

Eclipta

865

PricklySida

484

PalmerAmaranth

348

Sicklepod

241

Goosegrass

214

CutleafGroundcherry

116

dataset.info: none set.

Parsing decisions:

  • The dataset ships two parallel annotation formats: raw VIA JSON (annotation_VGG_json/) and a pre-converted YOLO-format txt (annotation_YOLO_txt/). This FiftyOne dataset’s ground_truth field was built directly from the raw VIA JSON, which sums to exactly 9,370 boxes and matches the paper/Zenodo card precisely. The shipped YOLO txt files sum to 9,388 boxes (18 extra, unexplained boxes across 14 files) and were not used — see Discrepancies below.

  • VIA JSON quirk handled during parsing: the regions key is a single dict when an image has exactly one box, and a list of dicts when it has 2+ boxes; both cases are normalized into a list before conversion.

  • Box coordinates were converted from VIA’s absolute-pixel, top-left x, y, width, height to FiftyOne’s relative [x, y, w, h] in [0, 1], using each image’s actual (variable) resolution rather than any fixed size.

  • Class labels were read from region_attributes.CottonWeed, a single-key dict per region (VIA checkbox-group attribute) — no multi-label or unchecked/false values were found in this release.

Dataset Creation#

Curation Rationale#

Weeds are a major threat to cotton production, and overreliance on herbicides for control has accelerated herbicide resistance and raised environmental, food-safety, and health concerns. The curators created CottonWeedDet12 to support development of robust, data-driven in-crop weed identification and detection systems (particularly YOLO-family object detectors) toward integrated, sustainable weed management for cotton and potentially other crops.

Source Data#

Data Collection and Processing#

Images were acquired with smartphones and hand-held digital cameras under natural field light conditions, at varied weed growth stages, in cotton fields at Michigan State University MEFAS field sites, primarily between June and September 2021 (per the paper/Zenodo card). Devices identified from filenames in this release include iPhone SE, iPhone 11 Pro, Nikon D3300, Canon EOS4000D, and a Huawei “YALAL00” model, at multiple resolutions (no single fixed image size).

Who are the source data producers?#

Field data collection was carried out by the research team at Michigan State University (Yuzhen Lu’s group), as described in the YOLOWeeds paper. Specific individual photographer identities beyond initials embedded in filenames are not documented.

Annotations#

Annotation process#

Images were manually annotated with rectangular bounding boxes for each visible weed instance using the VGG Image Annotator (VIA) v2.10. Each box was assigned exactly one of the 12 weed-class labels via a VIA checkbox-group attribute (CottonWeed).

Who are the annotators?#

“Qualified personnel” trained for weed identification, per the dataset’s Zenodo description. Specific annotator identities are not documented.

Personal and Sensitive Information#

None. The dataset contains only photographs of weeds and cotton-field vegetation; no personal or sensitive information is present.

Citation#

BibTeX:

@article{dang2023yoloweeds,
  title   = {YOLOWeeds: A novel benchmark of YOLO object detectors for multi-class weed detection in cotton production systems},
  author  = {Dang, Fengying and Chen, Dong and Lu, Yuzhen and Li, Zhaojian},
  journal = {Computers and Electronics in Agriculture},
  volume  = {205},
  pages   = {107655},
  year    = {2023},
  doi     = {10.1016/j.compag.2023.107655}
}

@dataset{lu2023cottonweeddet12,
  title     = {CottonWeedDet12: a 12-class weed dataset of cotton production systems for benchmarking AI models for weed detection},
  author    = {Lu, Yuzhen},
  year      = {2023},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.7535814}
}

APA:

Dang, F., Chen, D., Lu, Y., & Li, Z. (2023). YOLOWeeds: A novel benchmark of YOLO object detectors for multi-class weed detection in cotton production systems. Computers and Electronics in Agriculture, 205, 107655. https://doi.org/10.1016/j.compag.2023.107655

More Information#

The original Zenodo release ships two parallel annotation formats (raw VIA JSON and a pre-converted YOLO txt) that disagree by 18 boxes across 14 files; this FiftyOne dataset uses the raw VIA JSON as ground truth since it matches the published paper/card statistics exactly (9,370 boxes). See the companion benchmarking code at https://github.com/DongChen06/DCW for the original conversion scripts and YOLO detector training/evaluation pipeline used in the paper.

Dataset Card Authors#

Harpreet Sahota