Camponotus Fellah Trophallaxis Detection Dataset

Camponotus Fellah Trophallaxis Detection Dataset (camponotus-fellah-trophallaxis)

This dataset supports two-class object detection on Camponotus fellah in video-derived imagery related to trophallaxis (laboratory and in-situ sequences). Detection classes are normal (id 0) and trophallaxis (id 1). They correspond to CVAT state labels: non-trophallaxis behavior vs trophallaxis, exported from annotations that typically use a single box type with a state attribute (normal | trophallaxis). The string trackidmajor in folder names refers to the train/validation/test split rule, not to a third class: unique track_id values (identity tracks from CVAT) are first assigned to splits by ratio; then each image is assigned to the split of the most frequent track_id among its boxes (majority track). That reduces arbitrary scattering of consecutive frames across splits; minority tracks in the same frame can still fall under a different split than the image-level assignment (a known limitation of image-level evaluation with overlapping tracks).

The archive contains 2,511 images in a fixed split: train 1,772 / validation 694 / test 45. Images are duplicated under raw/ and prepared/ by design, so each export format can be used standalone without editing cross-folder paths.

Contents:
(1) raw/ — original full acquisition tree (as exported/collected).
(2) prepared/yolo_trackidmajor/ — Ultralytics YOLO layout: dataset.yaml (with portable path: .), images/{train,val,test}/, and normalized YOLO label .txt files.
(3) prepared/rfdetr_coco_trackidmajor/ — RF-DETR / Roboflow-style train/ and valid/ image folders with COCO JSON (train/_annotations.coco.json, valid/_annotations.coco.json); there is no separate test image folder here—held-out test images are defined via the COCO/YOLO exports in (2) and (4).
(4) prepared/coco_annotations_trackidmajor/ — COCO instances_train.json, instances_val.json, instances_test.json defining the full three-way split and category names.

License: MIT — see README.md in the archive for the full license text and copyright line.
Documentation: README.md (overview and usage), SOURCE_VERSIONS.txt (provenance and split counts), CHECKSUMS_SHA256.txt (SHA-256 for key metadata files such as README, dataset.yaml, manifests, and COCO JSON not per-image digests).
Related: Qualitative side-by-side demo Hugging Face Space: https://huggingface.co/spaces/dmytro-kushnir/ant-trophallaxis-rfdetr-vs-yolo26 = useful for visual comparison;

Keywords : Camponotus fellah, trophallaxis, object detection, COCO, YOLO, RF-DETR, CVAT, tracking, computer vision, MIT License.