dmytro-kushnir/camponotus-fellah-trophallaxis

---
pretty_name: Camponotus Fellah Trophallaxis Detection Dataset
license: mit
task_categories:
- object-detection
configs:
- config_name: examples_imagefolder
drop_labels: true
data_files:
- split: train
path:
- examples/train/*.jpg
- examples/train/*.jpeg
- examples/train/*.png
- examples/train/*.webp
- split: validation
path:
- examples/validation/*.jpg
- examples/validation/*.jpeg
- examples/validation/*.png
- examples/validation/*.webp
- split: test
path:
- examples/test/*.jpg
- examples/test/*.jpeg
- examples/test/*.png
- examples/test/*.webp
language:
- en
tags:
- camponotus-fellah
- ants
- trophallaxis
- computer-vision
- yolo
- rfdetr
size_categories:
- 1K<n<10K
---

# Camponotus Fellah Trophallaxis Detection Dataset

`camponotus-fellah-trophallaxis`

Two-class object detection dataset for *Camponotus fellah* behavior imagery.

- Class `0`: `normal` (non-trophallaxis state)
- Class `1`: `trophallaxis`

Annotations are exported from CVAT workflows where boxes can carry a behavioral `state` attribute (`normal` / `trophallaxis`).

## What is in this repository

- `prepared/yolo_trackidmajor/` - YOLO layout (`dataset.yaml`, `images/{train,val,test}`, labels)
- `prepared/rfdetr_coco_trackidmajor/` - RF-DETR/Roboflow-style layout (`train`, `valid`, `_annotations.coco.json`)
- `prepared/coco_annotations_trackidmajor/` - canonical COCO JSON split files (`instances_train.json`, `instances_val.json`, `instances_test.json`)
- `examples/` - curated sample images + `examples_manifest.json` for Space auto-preload
- `SOURCE_VERSIONS.txt` and `CHECKSUMS_SHA256.txt` - provenance and key-file checksums

Dataset Viewer is configured for `examples/` image previews. COCO JSON assets are included for download/use in training and evaluation pipelines.

## Splits

- Train: **1772** images
- Validation: **694** images
- Test: **45** images
- Total: **2511** images

## Track-majority split (`trackidmajor`)

`trackidmajor` indicates the split heuristic:

1. unique `track_id` values are assigned to train/val/test by ratio;
2. each image is assigned to the split of the majority `track_id` among its annotations.

This improves temporal consistency versus random per-image splitting, but minority tracks in an image can still belong to another split.

## Usage

### YOLO

Use `prepared/yolo_trackidmajor/dataset.yaml` directly.

### COCO / RF-DETR

Use:

- `prepared/coco_annotations_trackidmajor/instances_{train,val,test}.json`
- or RF-DETR-style `prepared/rfdetr_coco_trackidmajor/train|valid` + `_annotations.coco.json`

## Related resources

- Qualitative demo Space: https://huggingface.co/spaces/dmytro-kushnir/ant-trophallaxis-rfdetr-vs-yolo26
- Mendeley archival release: add your DOI/URL here
- Companion pipeline repo: `small-object-detection-benchmark`

## Citation

Please cite the dataset as:

> Camponotus Fellah Trophallaxis Detection Dataset (`camponotus-fellah-trophallaxis`), 2026.

Add DOI once assigned.

## License

MIT License (see `LICENSE`).