license: cc-by-4.0
task_categories:
- visual-question-answering
- image-text-to-text
tags:
- agriculture
- plant-pathology
- vqa
- vision-language
pretty_name: PlantMicro
configs:
- config_name: train
data_files:
- split: train
path: train-0000-of-0001.parquet
dataset_info:
config_name: train
splits:
- name: train
num_examples: 9718
size_categories:
- 1K<n<10K
PlantMicro
PlantMicro is a benchmark of microscopic plant imagery for evaluating vision-language models (VLMs) on fine-grained biological recognition at the cellular and subcellular level. Existing VLM benchmarks focus almost entirely on macroscopic, field-level plant imagery; PlantMicro was built to close that gap by evaluating models on microscopy, where domain-specific terminology and structures they were not trained for dominate.
The benchmark spans the domains of mycology, nematology, botany, and cell biology, across three imaging modalities (light, fluorescence, and electron microscopy) and 25 host species, sourced from tomato spores, PODB plant organelles, potato tubers, grape fungal spores, cucumber pathogens, plant-parasitic nematodes, and pollen imagery.
Paper: Benchmarking Vision-Language Models for Microscopic Plant Image Understanding (arXiv:2606.22497) Authors: Tianqi Wei, Xin Yu, Zhi Chen, Scott Chapman, Zi Huang Original repository: github.com/tqwei05/PlantMicro
About the paper
Construction. Images were gathered from peer-reviewed publications and institutional repositories across mycology, nematology, botany, and cell biology, prioritizing openly available data (with author permission requested where needed). Source images in heterogeneous formats (TIFF, JPEG, PNG) with inconsistent naming were standardized to JPEG, assigned unique identifiers, and given structured JSON metadata. Domain experts independently verified each image's attributes (biological domain, imaging modality, stain, host species), excluding inconsistent samples, and all 9,718 VQA pairs underwent expert review by biologists specializing in plant pathology and microscopy.
Tasks. Ten evaluation tasks span three levels: image-level (modality identification, domain classification, resin identification, stain/dye recognition), biological-level (host identification, pathogen classification, organelle detection, organ prediction), and quantitative (target counting, object localization/detection).
Models evaluated. Closed-source: GPT-4o mini, GPT-5 mini, GPT-5, Gemini 2.5-Flash-lite, Gemini 2.5-Flash, Gemini 2.5-Pro. Open-source: CLIP-ViT/32, CLIP-ViT/16, LLaVA-7B, LLaVA-13B, LLaVA-Next-7B, LLaVA-Next-13B, QwenVLM-7B.
Key findings. Models handle coarse, global visual cues well (modality identification: >95% for the best model) but struggle badly on fine-grained biological recognition: pathogen classification tops out at 34.93% (GPT-5), only modestly above a 24.95% random-guessing baseline, and host identification averages ~30% across models, also close to chance. Closed-source models outperform open-source ones by 10+ percentage points on average. Error analysis attributes most failures to knowledge deficiency rather than perception error — models can see the relevant structures but don't know what they mean.
Dataset Size
| Config | Rows (QA pairs) | Unique Images | Image Data |
|---|---|---|---|
| train | 9,718 | 5,369 | 0.54GB (1 shard) |
The source dataset has 5,410 images, each with one or more associated questions (9,718 VQA pairs total). Standardization flattens each (image, question) pair into its own row, so the row count here (9,718) is higher than the source image count (5,410) — it matches the VQA pair count, not the image count. The 5,369 unique images (vs. 5,410 source entries) reflects a small number of images that are byte-identical duplicates, deduplicated via SHA256 hash.
Question Types
| Question type | Mechanism | Count |
|---|---|---|
| modality, class, domain, specie, stain, organell, resin, organ | multiple-choice (options) |
8,030 |
| count_c | multiple-choice (choices + correct_choice_idx) |
605 |
| detection | open-ended bounding-box coordinates | 1,083 |
Layout
train-0000-of-0001.parquet # 9,718 rows
images/
image-train-000-of-001.zip # image shard (ZIP_STORED, uncompressed)
path_to_shard.parquet # maps image SHA256 hash -> shard_file
Images are content-addressed by SHA256 hash rather than by their original file path, and are stored uncompressed in zip shards for random access via zipfile. Each image shard stays under 5GB.
Schema
| Column | Type | Description |
|---|---|---|
images |
list<struct<bytes, path>> |
bytes is null; path is the SHA256 hash used to look up the image in images/path_to_shard.parquet |
id |
string |
plantmicro_train_<n> |
messages |
chat-style list | user turn: image + question (multiple-choice questions include lettered options inline); assistant turn: the answer, letter-prefixed for multiple-choice questions |
raw_metadata |
JSON string | dataset, question_type, options, choices, correct_choice_idx, image_path (original source path) |
Each row is one (image, question) pair; a single source image can appear across multiple rows since PlantMicro asks multiple questions per image.
Usage
Recommended, via the AgML python library:
from agml import loadImageTextToTextDataset
ds = loadImageTextToTextDataset("Project-AgML/PlantMicro")
Citation
@article{wei2026plantmicro,
title={Benchmarking Vision-Language Models for Microscopic Plant Image Understanding},
author={Wei, Tianqi and Yu, Xin and Chen, Zhi and Chapman, Scott and Huang, Zi},
journal={arXiv preprint arXiv:2606.22497},
year={2026}
}
This dataset is indexed and structured on https://project-agml.github.io/ as part of the AgML python library. This dataset was reformatted from its original format to match HuggingFace's Imagefolder standards but requires an external module (agml) that processes and returns a HF Dataset object faster than HF module functions.