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.

Dataset Card for PlantWild#

image/png

This is a FiftyOne dataset with 30030 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/PlantWild")

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

Dataset Details#

Dataset Description#

PlantWild is a large-scale, in-the-wild, multimodal plant disease recognition dataset. Unlike lab-captured datasets (e.g. PlantVillage), PlantWild’s images are crowdsourced from internet search engines, giving them realistic complex backgrounds, viewpoints, and lighting. Each disease class is additionally paired with 50 GPT-3.5-generated descriptive text prompts, making the dataset a testbed for multimodal (image + text) disease recognition, including CLIP-style zero-shot and few-shot classification.

Two releases exist and are both included in this FiftyOne dataset:

v1 (the version described in the paper): 18,542 images across 89 classes (56 diseased + 33 healthy-plant classes), with an official train/val/test split, per-image source URLs, and the 50 text prompts per class.
v2: a later, expert-refined release with 11,488 images across 115 disease-only classes (healthy-plant classes were dropped, many new fine-grained disease classes were added). No split/URL/prompt metadata ships with v2.
Curated by: Tianqi Wei, Zhi Chen, Zi Huang, Xin Yu. The source repository owner (uqtwei2) and one of the official download mirrors (UQRDM, the University of Queensland’s Research Data Manager) indicate the work originates from the University of Queensland.
Funded by: Grains Research and Development Corporation
Shared by: Original dataset shared on the Hugging Face Hub by uqtwei2. This FiftyOne-formatted version was parsed and shared by harpreetsahota.
Language(s): English (all 50 text prompts per class are in English; Chinese was used only as an auxiliary keyword language during image collection via Baidu Images, and is not present anywhere in the released data).
License: cc-by-nc-nd-4.0 (Attribution-NonCommercial-NoDerivatives 4.0). This FiftyOne mirror repackages the same images and labels (no creative modification) with full attribution; redistribution/use should respect the NonCommercial and NoDerivatives terms of the original license.

Dataset Sources#

Repository: https://huggingface.co/datasets/uqtwei2/PlantWild (source data), https://github.com/tqwei05/MVPDR (official baseline code and data loaders)
Paper: Wei, T., Chen, Z., Huang, Z., & Yu, X. (2024). Benchmarking In-the-Wild Multimodal Plant Disease Recognition and A Versatile Baseline. ACM International Conference on Multimedia (ACM MM 2024).
Demo: https://tqwei05.github.io/PlantWild/ (project homepage)

Uses#

Direct Use#

Benchmarking in-the-wild plant disease image classification (fully-supervised, few-shot, and zero-shot/training-free settings).
Multimodal research combining visual and textual (CLIP-style) representations, e.g. building per-class visual/textual prototypes as in the paper’s MVPDR baseline.
Studying long-tailed, fine-grained visual classification with small inter-class discrepancy and large intra-class variance (the paper’s central motivation).
Cross-version taxonomy comparison and healthy-vs-diseased classification research (v1 only, since v2 dropped healthy classes).

Out-of-Scope Use#

Commercial use or redistribution of modified/derivative versions — disallowed by the source cc-by-nc-nd-4.0 license.
Real-world agricultural diagnosis — this dataset is intended for ML research/benchmarking; it is not a validated diagnostic tool and should not be used as the sole basis for real crop treatment decisions.
Pixel-level disease localization — PlantWild provides only image-level classification labels, no segmentation masks. A companion dataset, PlantSeg, provides segmentation annotations for diseased regions and is better suited for that task.
Naively balanced training — class counts are long-tailed (7–589 images/class across both versions combined), so uniform sampling/evaluation should account for this.

Dataset Structure#

This is a flat image FiftyOne dataset (no groups/frames) with 30,030 samples, combining PlantWild v1 and v2 into a single dataset distinguished by a dataset_version field (rather than two separate FiftyOne datasets), so that classes shared across both taxonomies can be queried and compared directly.

Sample fields#

Field	FiftyOne type	Description
`filepath`	`StringField`	Path to the image file
`ground_truth`	`Classification`	Disease (or healthy-leaf) label. 146 distinct classes total across both versions (89 in v1, 115 in v2, 58 shared)
`dataset_version`	`StringField`	`"v1"` or `"v2"` — which source archive the sample came from
`split`	`StringField`	`"train"` / `"val"` / `"test"` for v1 (the paper’s official 70/10/20 per-class split, verbatim from `trainval.txt`); `None` for v2, which ships no official split
`source_url`	`StringField`	Original web URL the image was scraped from, verbatim from v1’s `url_record.json` (v1 only; `None` for v2)
`disease_prompts`	`ListField(StringField)`	The 50 GPT-3.5-generated descriptive text prompts for the sample’s class, verbatim from `plantwild_prompts.json`. Present for all v1 classes and for the 58 v2 classes whose name matches a v1 class; absent (`None`) for the 57 disease classes newly introduced in v2
`tags`	`ListField(StringField)`	`"v1"`/`"v2"` plus the split name (for v1), for quick filtering in the App
`metadata`	`ImageMetadata`	`width`, `height`, `size_bytes`, `mime_type`, computed via `compute_metadata()`

Image dimensions range from 36×39px to 8256×5504px and file sizes from 1.4KB to 11.2MB, reflecting the “in-the-wild” crowdsourced collection process.

`dataset.info`#

{
    "source": "https://huggingface.co/datasets/uqtwei2/PlantWild",
    "paper": "Benchmarking In-the-Wild Multimodal Plant Disease Recognition and A Versatile Baseline (ACM MM 2024) - https://arxiv.org/abs/2408.03120",
    "license": "cc-by-nc-nd-4.0",
    "versions": {
        "v1": "18,542 images, 89 classes (incl. healthy-leaf classes), official train/val/test split, source URLs, 50 GPT-generated text prompts per class",
        "v2": "11,488 images, 115 disease-only classes, expert-refined, no split/url metadata",
    },
}

Indexes#

Single-field indexes on ground_truth.label, dataset_version, split, tags, metadata.size_bytes, metadata.width, metadata.height; compound indexes on (dataset_version, split) and (dataset_version, ground_truth.label).

Saved views#

View	Samples	Description
`v1-full`	18,542	All PlantWild v1 samples
`v2-full`	11,488	All PlantWild v2 samples
`v1-train`	13,045	Official v1 training split
`v1-val`	1,820	Official v1 validation split
`v1-test`	3,677	Official v1 test split
`shared-classes`	16,864	Samples whose class name exists in both v1 and v2 taxonomies (58 classes)
`v2-novel-diseases`	5,209	v2 samples from the 57 disease classes not present in v1
`healthy-leaves`	7,840	Non-diseased `"<plant> leaf"` reference images (v1 only, 30 classes)
`diseased-only`	22,190	All samples with an actual disease label
`rare-classes`	98	Classes with fewer than 20 images (7 classes) — long-tail/few-shot candidates
`likely-duplicate-filenames`	673	Filenames containing `"Copy"`/`"Copy (n)"` — Windows-style duplicate naming, flagged for dedup review
`low-resolution`	181	Images with width or height under 100px
`tiny-filesize`	2	Images under 2KB, possibly corrupt/blank

Parsing decisions#

Combined, not split, into one dataset. v1 and v2 are stored together with a dataset_version tag/field rather than as two separate FiftyOne datasets, so overlapping classes (58 of them) can be compared directly across versions.
Split semantics validated against the authors’ own code. The official MVPDR data loader (datasets/plantwild.py) encodes trainval.txt modes as 1=train, 2=val, 0=test(anything else)=test, and only regenerates a random split if the file is missing — since it already ships in the archive, the authors’ own pipeline (like ours) just parses it directly. We independently re-derived the per-class split-generation formula (train = ceil(ceil(0.8·n)·0.875), val = remainder of the 80%, test = the other 20%) and confirmed 0 mismatches across all 89 v1 classes against our parsed split field.
disease_prompts is class-level, not per-image, metadata. The source plantwild_prompts.json is keyed by class name, not by image, and the official data loader never references it — it’s consumed elsewhere in the MVPDR pipeline to build per-class CLIP textual prototypes. We attach the identical 50-prompt list to every sample of a class (rather than splitting one prompt per image) to stay faithful to this class-level semantic, accepting some storage duplication in exchange for simpler per-sample querying in the App.
Class names come from folder names, not classes.txt. This matches the official loader’s own approach (self.classes = sorted(os.listdir(image_dir))), avoiding any risk of an id-to-name mismatch.
v2 has no official split/url/prompt metadata because no plantwild_v2.py loader or accompanying files were ever published — split and source_url are None for all v2 samples, and disease_prompts is only populated for the 58 v2 classes whose names happen to match a v1 class.
Data-quality issues were flagged as views, not removed. 673 images have “Copy”-style filenames suggesting literal duplicates, and 181 images are under 100px — both are preserved in the dataset but exposed as saved views for downstream review/dedup rather than silently dropped.

Dataset Creation#

Curation Rationale#

The paper’s authors note that existing plant disease datasets were insufficient for real-world use: PlantVillage (54,309 images, 38 classes) is lab-captured only, lacking real-world complexity, while the existing in-the-wild alternative, PlantDoc, has just 2,598 images across 27 classes — too small in practice. The authors also observed that in-the-wild plant disease images exhibit small inter-class discrepancy (different diseases can look visually similar) and large intra-class variance (the same disease can look very different across images), motivating the addition of GPT-3.5-generated textual descriptions as a second modality to help disambiguate visually similar diseases.

Source Data#

Data Collection and Processing#

Images were crowdsourced from Google Images, Ecosia, and Baidu Images (the latter added specifically because China’s large and varied farmland made Chinese-language sources valuable), using each disease’s common name in both English and Chinese as search keywords. Collection yielded more than 50,000 raw images, which were then filtered down to the final 18,542-image, 89-class v1 release (56 diseased + 33 healthy classes, 44–589 images per class). PlantWild v2 (115 disease-only classes, 11,488 images) is a later expert-refined release, described on the project homepage as adding new disease classes based on expert agricultural feedback; it postdates the paper and is not described in it.

Who are the source data producers?#

Images were sourced from public internet image search engines (Google Images, Ecosia, Baidu Images), i.e. crowdsourced from a large number of original, uncredited photographers and websites. Per-image source URLs are preserved in v1’s source_url field for provenance tracing back to the original hosting page.

Annotations#

Annotation process#

Five annotators filtered the crowdsourced images against reference disease exemplar images from the University of Minnesota (UMN) extension, removing irrelevant or low-quality images. Each remaining image’s class label was cross-validated by at least two annotators; when annotators disagreed, a third expert reviewed and resolved the label. Descriptive text prompts (50 per class) were generated using GPT-3.5, informed by Wikipedia descriptions of each disease.

Who are the annotators?#

Five annotators plus at least one additional domain expert for dispute resolution, per the paper. Specific identities/qualifications are not disclosed beyond this.

Personal and Sensitive Information#

None identified. The dataset consists exclusively of plant/plant-disease imagery and disease-descriptive text; no personal or sensitive information is present.

Citation#

BibTeX:

@inproceedings{MVPDR,
  title={Benchmarking In-the-Wild Multimodal Plant Disease Recognition and A Versatile Baseline},
  author={Wei, Tianqi and Chen, Zhi and Huang, Zi and Yu, Xin},
  booktitle={ACM International Conference of Multimedia},
  year={2024}
}

APA:

Wei, T., Chen, Z., Huang, Z., & Yu, X. (2024). Benchmarking in-the-wild multimodal plant disease recognition and a versatile baseline. In Proceedings of the 32nd ACM International Conference on Multimedia (ACM MM 2024).

More Information#

Project homepage: https://tqwei05.github.io/PlantWild/
Official baseline code (MVPDR) and data loaders: https://github.com/tqwei05/MVPDR
A companion dataset, PlantSeg, extends PlantWild with segmentation annotations for diseased regions (see the project homepage).
Source dataset (v1 + v2 raw archives): https://huggingface.co/datasets/uqtwei2/PlantWild

Dataset Card Authors#

Parsed into FiftyOne format and dataset card authored by harpreetsahota, based on the original dataset and paper by Wei, Chen, Huang, and Yu.

Dataset Card Contact#

For questions about the original PlantWild data/labels, use the source dataset’s discussion page. For questions about this FiftyOne-formatted mirror, contact harpreetsahota.