---
annotations_creators:
- expert-generated
language: en
license: cc-by-sa-4.0
size_categories:
- 1K<n<10K
task_categories:
- object-detection
- image-segmentation
task_ids:
- semantic-segmentation
- instance-segmentation
pretty_name: PhenoBench
tags:
- agriculture
- agtech
- crops
- drone
- fiftyone
- image
- image-segmentation
- instance-segmentation
- object-detection
- phenotyping
- semantic-segmentation
- sugar-beet
- uav
- weeds
---

![pheno-bench-demo](./pheno-bench-demo.gif)

# Dataset Card for PhenoBench

This is a [FiftyOne](https://github.com/voxel51/fiftyone) dataset with 2,179 samples.

The images and original annotations are from [PhenoBench](https://www.phenobench.org/),
a large UAV image dataset for semantic image interpretation in the agricultural
domain (sugar beet crops and weeds), introduced by Weyler et al. in
[*PhenoBench — A Large Dataset and Benchmarks for Semantic Image Interpretation
in the Agricultural Domain*](https://arxiv.org/abs/2306.04557) (IEEE TPAMI,
2024). This release packages a subset of PhenoBench as a FiftyOne dataset and
adds object-detection predictions and embeddings produced by Voxel51.

## What's in this dataset

- **2,179 samples** (train: 1,407 / val: 386 / test: 386), each a 1024×1024 RGB
  UAV image of a sugar beet field
- **Original PhenoBench annotations** (from phenobench.org):
  - `semantics` — semantic segmentation: background, crop, weed, partial-crop,
    partial-weed
  - `plant_instances` — per-plant instance segmentation
  - `leaf_instances` — per-leaf instance segmentation
  - `plant_visibility`, `leaf_visibility` — visibility heatmaps
- **Voxel51-added detections and embeddings**:
  - `yolo11n`, `yolo11l` — YOLO11 (nano and large) object detection predictions
    for plants, with per-detection true-positive / false-positive /
    false-negative matches against the ground-truth instance masks
  - Brain runs in `brain/`: CLIP and DINOv2 embeddings + similarity indexes for
    sample-level and patch-level visual search

For the full original PhenoBench dataset (train/val/test = 1,407 / 772 / 693)
and the canonical annotation specification, see [phenobench.org](https://www.phenobench.org/).

## License

CC BY-SA 4.0, inherited from the upstream PhenoBench release.

## Installation

```bash
pip install -U fiftyone
```

## Usage

```python
import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub

# Load the dataset
dataset = load_from_hub("Voxel51/PhenoBench")

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

## Citation

If you use this dataset, please cite the original PhenoBench paper:

```bibtex
@article{weyler2024phenobench,
  title={PhenoBench: A Large Dataset and Benchmarks for Semantic Image Interpretation in the Agricultural Domain},
  author={Weyler, Jan and Magistri, Federico and Marks, Elias and Chong, Yue Linn and Sodano, Matteo and Roggiolani, Gianmarco and Chebrolu, Nived and Stachniss, Cyrill and Behley, Jens},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  year={2024},
  url={https://arxiv.org/abs/2306.04557}
}
```

Please refer to [phenobench.org](https://www.phenobench.org/) for the
authoritative citation and licensing terms.