---
license: mit
pretty_name: BeeMachine Partwhole Dataset
task_categories:
- image-segmentation
- image-classification
tags:
- biology
- entomology
- bumble-bee
- fine-grained
- part-segmentation
- citizen-science
size_categories:
- 1K<n<10K
---

# BeeMachine Partwhole Dataset

Pixel-level anatomical part annotations for bumble bee photographs used by the BeeMachine identification service and by the IAAI-27 evaluation of anatomy-guided learning strategies.

## Summary

| Property | Value |
|---|---|
| Images | 7,716 |
| Species | 160 |
| Part classes | background, abdomen, head, thorax |
| Mask format | integer PNG (`*_m.png`), same stem as the RGB image |
| Labels | `species_labels.csv` (`images`, `species`) |

## Directory layout

Files in this repository:

```
README.md
species_labels.csv      # image filename → species
annotations.coco.json   # COCO-format segmentation export
images.tar.gz           # RGB photographs (flat directory `images/`)
masks.tar.gz            # integer part masks (flat directory `masks/`)
```

After download, unpack next to the CSV:

```bash
tar -xzf images.tar.gz
tar -xzf masks.tar.gz
```

This yields:

```
images/                 # RGB photographs (flat)
masks/                  # integer part masks, one {stem}_m.png per image
species_labels.csv
annotations.coco.json
```


Mask class IDs:

| ID | Label |
|---:|---|
| 0 | background |
| 1 | abdomen |
| 2 | head |
| 3 | thorax |

## Intended use

Train and evaluate part segmenters that supply anatomical masks to species classifiers. In the associated paper, this corpus is the Stage A segmentation set only; species classification is trained on a separate, larger originals-only BeeMachine classification corpus.

**Note:** The full BeeMachine classification dataset is very large and is not hosted here; it will be available upon request.

## Splits

The paper uses a frozen species-stratified 75% / 15% / 10% train / validation / test split (seed 42). Those fold filenames are distributed with the supplementary code under `outputs/beemachine/frozen_splits/`. Train-only six-fold geometric augmentation (original, flips, 90/180/270° rotations) is generated from the frozen training fold and is not shipped here.

## Source photographs

Images were collected from citizen-science and museum archives used by BeeMachine (including Bumble Bee Watch, iNaturalist, and BugGuide). Part masks were annotated for head, thorax, and abdomen. Copyright in the underlying photographs remains with the original rights holders; redistribution here is for non-commercial research.

## Citation

If you use this dataset, please cite the BeeMachine paper and the IAAI-27 evaluation that releases these part annotations:

```bibtex
@article{spiesman2021beemachine,
  author  = {Spiesman, Brian J. and Gratton, Claudio and Hatfield, Richard G. and Hsu, William H. and Jepsen, Sarina and McCornack, Brian and Patel, Krishna and Wang, Guanjie},
  title   = {Assessing the Potential for Deep Learning and Computer Vision to Identify Bumble Bee Species from Images},
  journal = {Scientific Reports},
  volume  = {11},
  pages   = {7580},
  year    = {2021},
  doi     = {10.1038/s41598-021-87210-1}
}
```

## License

MIT for the annotation files and packaging in this repository. Photograph copyrights remain with their original owners.