Sainath001/stomata-keypoint-benchmark-cvpr-agrivision-2026-Dataset

Stomata Keypoint Detection: Dataset

This repository contains a multi-domain dataset for stomata keypoint detection. It was created to evaluate how well keypoint detection models hold up under distribution shift across imaging environments, plant groups, species, and microscope systems.

The benchmark accompanies the following paper:

Towards Morphology Aware Stomata Keypoint Detection: Benchmarking Foundation Models Under Distribution Shift
Accepted at the CVPR 2026 AgriVision Workshop

• Paper: Coming soon
• Code: Coming soon on GitHub
• Models: stomata-keypoint-benchmark-cvpr-agrivision-2026

---

Overview

The benchmark includes 1 training split and 9 test splits. Together, they cover several types of distribution shift, including field-to-lab transfer, monocot-to-dicot transfer, species changes, and sensor changes.

Each stomata instance is annotated with 4 keypoints in COCO format:

• two polar tips along the stomatal length axis
• two lateral endpoints along the stomatal width axis

Dataset Summary

Split	Environment	Botanical Class	Microscope	Species	Images	Stomata
KP-Train	Field	Monocot	ProScope HR	Maize	344	12,503
T-MZLP	Field	Monocot	ProScope HR	Maize	86	3,188
T-MZA	Field	Monocot	ProScope HR	Maize	115	4,349
T-SBGH	Greenhouse	Dicot	ProScope HR	Soybean	13	524
T-MSAA	Field	Dicot	ProScope HR	Nokaidō	12	444
T-HR5MZ	Lab	Monocot	ProScope HR5	Maize	10	924
T-HR5WH	Greenhouse	Monocot	ProScope HR5	Wheat	6	153
T-TCMZ	Lab	Monocot	Toupcam S500-GS	Maize	25	270
T-NKBY	Lab	Monocot	Nikon DS-Fi3	Barley	15	175
T-TCAB	Lab	Dicot	Toupcam S500-GS	Arabidopsis	13	134

Total: 639 images and 22,664 annotated stomata

---

What This Benchmark Measures

The test splits were designed to probe robustness under several kinds of distribution shift relative to KP-Train.

Shift Type	Test Splits
Location shift	T-MZA
Plant-level shift	T-MZLP
Environment shift	T-SBGH, T-HR5MZ, T-TCMZ
Taxonomic shift	T-SBGH, T-MSAA, T-TCAB
Species shift	T-HR5WH, T-NKBY, T-MSAA
Sensor/system shift	T-HR5MZ, T-HR5WH, T-TCMZ, T-TCAB, T-NKBY

This setup makes the benchmark useful for testing not only in-domain accuracy, but also how well a model generalizes to new biological and imaging conditions.

---

Annotation Format

Annotations are provided in COCO keypoint format.

Each stomata instance includes:

• a bounding box: [x, y, width, height]
• four keypoints: [x0, y0, v0, x1, y1, v1, x2, y2, v2, x3, y3, v3]

Keypoint Definition

• p0, p1: polar tips along the stomatal length axis
• p2, p3: lateral endpoints along the stomatal width axis

Annotation Source

For each dataset split:

• annotations/ contains the original annotation files
• coco.json contains the COCO-format annotations

The original annotations inside annotations/ include Pascal VOC XML files, and the COCO annotations were generated from those annotations using RectLabel Pro.

---

Repository Structure

datasets/
├── KP-Train/
│   ├── annotations/
│   ├── images/
│   └── coco.json
├── T-MZLP/
│   ├── annotations/
│   ├── images/
│   └── coco.json
├── T-MZA/
│   ├── annotations/
│   ├── images/
│   └── coco.json
├── T-SBGH/
│   ├── annotations/
│   ├── images/
│   └── coco.json
├── T-MSAA/
│   ├── annotations/
│   ├── images/
│   └── coco.json
├── T-HR5MZ/
│   ├── annotations/
│   └── coco.json
├── T-HR5WH/
│   ├── annotations/
│   └── coco.json
├── T-TCMZ/
│   ├── annotations/
│   └── coco.json
├── T-NKBY/
│   ├── annotations/
│   └── coco.json
└── T-TCAB/
    ├── annotations/
    └── coco.json

---

External Data Sources

Some test splits contain images derived from previously published dataset. These sources are listed below for proper attribution.

Splits T-HR5MZ, T-TCMZ, T-TCAB, T-HR5WH, and T-NKBY include images originating from the following works:

1. Xiaohui Yang, Jiahui Wang, Fan Li, Chenglong Zhou, Minghui Wu, Chen Zheng, Lijun Yang, Zhi Li, Yong Li, Siyi Guo, et al. RotatedStomataNet: A Deep Rotated Object Detection Network for Directional Stomata Phenotype Analysis. Plant Cell Reports, 43(5):126, 2024. Repository: https://github.com/AITAhenu/RotatedStomataNet/tree/main/test-images

2. Phetdalaphone Pathoumthong, Zhen Zhang, Stuart J. Roy, and Abdeljalil El Habti. Rapid Non-destructive Method to Phenotype Stomatal Traits. Plant Methods, 19(1):36, 2023. Repository: https://github.com/rapidmethodstomata/rapidmethodstomata

3. Na Sai, James Paul Bockman, Hao Chen, Nathan Watson Haigh, Bo Xu, Xueying Feng, Adriane Piechatzek, Chunhua Shen, and Matthew Gilliham. StomataAI: An Efficient and User-friendly Tool for Measurement of Stomatal Pores and Density Using Deep Computer Vision. New Phytologist, 238(2):904–915, 2023. Repository: https://github.com/xdynames/sai-app

The keypoint annotations for these benchmark splits were newly created by Sainath Reddy Gummi.

If you redistribute this benchmark, please source the external images from above citations and github links.

---

License

Component	License
Images and annotations for internal splits	CC BY-NC 4.0
Keypoint annotations created for external-image splits	CC BY-NC 4.0

---

Citation

If you use this dataset and annotations, please cite:

@inproceedings{gummi2026stomata,
  author    = {Gummi, S. R. and Pack, C. and Zhang, H. K. and Solanki, S. and Chang, Y.},
  title     = {Towards Morphology Aware Stomata Keypoint Detection: Benchmarking Foundation Models Under Distribution Shift},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops},
  year      = {2026},
  note      = {Accepted}
}

---

Contact

Sainath Reddy Gummi South Dakota State University Email: gummisainath@gmail.com