Dataset Card for Arecibo C3 Bee Feeder Videos (AC3_BeeVids)

Raw high-resolution video recordings of paint-marked honey bees (Apis mellifera) at artificial feeder stations, collected for the Arecibo C3 (AC3) Center citizen-science and behavioral-ecology project in Puerto Rico. The videos are packaged in WebDataset (.tar) shards and split into three configurations, one per school feeder site: American-Military-Academy, Monserrate-León-de-Irizarry, and Pedro-Falú.

This collection is the raw upstream source from which the downstream re-identification dataset megretlab/red_bee_reID (cropped, masked and annotated bee images) was derived, and it underpins the WACV 2026 paper One-Shot Fine-Grained Re-Identification of Paint Marked Honey Bees using Vision Foundation Models. It is also a companion to the AC3 foraging-log dataset published on Zenodo (10.5281/zenodo.20736946).

Dataset Details

Dataset Description

The AC3 project tracks honey-bee foraging behavior at artificial feeders installed at partner schools in Puerto Rico. At each feeder, foragers are individually marked with paint dots and recorded with overhead machine-vision cameras as they arrive at and depart from the feeder entrance ramp. These raw videos support studies of foraging dynamics (bee-lining / "bee hunting", arrival–departure timing, navigation) as well as computer-vision research on detection, tracking, and fine-grained re-identification of individual insects.

• Curated by: Megret Lab and the AC3 team — University of Puerto Rico, Río Piedras, in partnership with participating schools (e.g., Escuela Pedro Falú Orellano, Río Grande).
• Funded by: U.S. National Science Foundation, Arecibo C3 (AC3) Center, awards #2321759, #2321760, and #2321761 (project lead Dr. José L. Agosto Rivera). Related computer-vision work was supported by NSF #2318597 (CyIndiBee) and used the UPR High-Performance Computing facility (NIH/NIGMS award 5P20GM103475).
• Shared by: megretlab on the Hugging Face Hub.
• Language(s): English (eng) for all metadata and documentation.
• License: Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).

Dataset Sources

• Repository: https://huggingface.co/datasets/megretlab/AC3_BeeVids
• Derived re-ID dataset: https://huggingface.co/datasets/megretlab/red_bee_reID
• Foraging-log companion (Zenodo): https://doi.org/10.5281/zenodo.20736946
• Paper: One-Shot Fine-Grained Re-Identification of Paint Marked Honey Bees using Vision Foundation Models, WACV 2026 — https://openaccess.thecvf.com/content/WACV2026/papers/Meyers_One-Shot_Fine-Grained_Re-Identification_of_Paint_Marked_Honey_Bees_using_Vision_WACV_2026_paper.pdf
• Code: https://github.com/megretlab/bee_reid_dinov3

Uses

Direct Use

• Honey-bee detection and multi-object tracking in top-down feeder video.
• Fine-grained re-identification of individual bees from paint-mark and natural appearance cues (the primary motivation for red_bee_reID).
• Behavioral-ecology analysis of foraging: arrival/departure timing, feeder visitation rates, and bee-lining/"bee hunting" studies.
• A raw-video benchmark for developing video processing, cropping, masking, and keypoint pipelines.

Out-of-Scope Use

• The CC BY-NC-ND 4.0 license prohibits commercial use and redistribution of modified/derivative versions of the videos. Derived works (crops, annotations, models) should be released separately under their own terms — see the derived dataset for an example.
• The data is from a single region (Puerto Rico), a small number of days, and three sites, so it is not intended as a representative sample of honey-bee behavior across populations, seasons, or geographies.

Dataset Structure

The dataset is organized into three configurations, one per feeder site, each backed by one or more WebDataset .tar shards:

American-Military-Academy/*.tar
Monserrate-León-de-Irizarry/*.tar
Pedro-Falú/*.tar

Data Instances

Each WebDataset shard is a .tar archive of samples. A sample corresponds to a video recording (or a clip thereof) keyed by a unique sample id, with the encoded video stream stored as the primary member and any associated metadata stored as a sidecar member sharing the same key. Loading is supported through the WebDataset API or the Hugging Face datasets viewer.

Note: the exact per-sample member keys (e.g. video file extension and metadata sidecar fields) should be confirmed against the published shards and refined here. [More Information Needed — exact WebDataset sample keys]

Data Fields

• Video stream: H.265/HEVC-encoded recording at 3840×2160 px, 17 fps (see Data Collection).
• Sample key / id: unique identifier per recording within a site.
• Sidecar metadata (if present): source site, recording session/date, and provenance fields. [More Information Needed — confirm exact field names]

Data Splits

Splits are organized by recording site (the three configs above) rather than by train/validation/test. Per-site shard counts and total video duration: [More Information Needed].

Dataset Creation

Curation Rationale

The videos were collected to study honey-bee foraging behavior at artificial feeders within the AC3 project and to provide raw material for computer-vision research on individual-bee tracking and re-identification. Marking foragers and filming them at the feeder entrance enables both behavioral quantification and the construction of identity-labeled image datasets.

Source Data

Data Collection and Processing

Recordings follow a shared AC3 protocol across the three feeder sites (documented in detail at the Pedro-Falú / Río Grande, Puerto Rico site, which is the source for the derived re-ID dataset):

• Setup: an overhead machine-vision camera films the entrance ramp of an artificial bee feeder.
• Camera: Basler a2A3840-45ucPRO, 3840×2160 px resolution at 17 fps, encoded as H.265/HEVC video.
• Marking: as many foragers as possible are marked with a single dot of paint on the thorax (red or green) so individuals can be distinguished across visits.
• Coverage (documented site): approximately 9 videos recorded over 3 days at the Pedro-Falú feeder in Río Grande. The American-Military-Academy and Monserrate-León-de-Irizarry sites follow the same protocol; their exact recording counts and dates are [More Information Needed].
• Packaging: raw recordings are organized per site and serialized into WebDataset .tar shards for distribution.

Downstream, these videos were processed into the red_bee_reID dataset by extracting per-bee crops (≈1.5× the average bee-skeleton length), masking the background with SAM2, rotating crops to a head–abdomen vertical alignment, and adding keypoint and identity annotations.

Who are the source data producers?

The AC3 team at the University of Puerto Rico, Río Piedras, and partner schools, including (from the companion Zenodo record) Lizbeth Alvarado Vargas, Ariana I. Rodríguez Flores, Luis Aparicio Mestra, Jesús Castro González, Jaime W. Abreu Ramos, Andrea Ortiz Cana, Amilcar Velez Flores, Jairo A. Ayala-Godoy, Tugrul Giray, and project leader José L. Agosto Rivera. Video collection and processing for the re-identification work were carried out by Luke Meyers, Josué A. Rodríguez-Cordero, and Rémi Mégret.

Annotations

This release contains raw video without per-frame labels. Identity, tracking, keypoint, and crop annotations were produced downstream and are distributed with the derived red_bee_reID dataset rather than here.

Annotation process

Not applicable to this raw-video release. See red_bee_reID for the downstream annotation process (manual identity assignment, automated tracking, keypoint estimation).

Who are the annotators?

Not applicable to this release; see the derived dataset.

Personal and Sensitive Information

The subjects are insects (honey bees); the dataset contains no personally identifiable human information. Recordings are framed on the feeder ramp; any incidental appearance of people (e.g., a hand placing marks) is minimal. If reviewers identify any incidental human imagery, it should be reported to the maintainers. [More Information Needed if any human subjects appear]

Bias, Risks, and Limitations

• Limited scope: three feeder sites in a single region (Puerto Rico) over a small number of days; not representative of honey-bee behavior across populations, seasons, or environments.
• Sampling bias: only foragers that could be caught and paint-marked are individually identifiable, and marking favors active, frequently visiting bees.
• Environmental variation: lighting, weather, and feeder traffic vary across sites and sessions.
• License constraints: the NoDerivatives clause limits remixing/redistribution of altered videos.

Recommendations

Users should treat results as site- and season-specific, report which configuration(s) were used, and release any derived annotations or models under separately stated licenses. Cross-site generalization should be validated explicitly.

Licensing Information

Released under CC BY-NC-ND 4.0 (Attribution, NonCommercial, NoDerivatives). Note that the derived re-identification dataset red_bee_reID is released under CC BY-NC 4.0; check each resource's terms before use.

Citation

If you use this dataset, please cite the associated paper:

BibTeX:

@inproceedings{meyers2026one,
  title={One-Shot Fine-Grained Re-Identification of Paint Marked Honey Bees using Vision Foundation Models},
  author={Meyers, Luke and Rodr{\'\i}guez-Cordero, Josu{\'e} A and M{\'e}gret, R{\'e}mi},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  pages={560--569},
  year={2026}
}

Please also cite the companion AC3 foraging-records dataset where relevant: DOI 10.5281/zenodo.20736946.

Acknowledgements

This work was supported by NSF award #2318597 (CyIndiBee). Data collection was possible through work supported by NSF award #2321760 under Dr. J. Agosto Rivera, with the Arecibo C3 (AC3) Center (NSF #2321759, #2321760, #2321761). Special thanks to L. Alvarado Vargas, A. Rodríguez, and M. Geria. This work used the UPR High-Performance Computing facility, supported by NIH/NIGMS award 5P20GM103475.

Dataset Card Authors

Luke Meyers, Josué A. Rodríguez-Cordero, and Rémi Mégret.

Dataset Card Contact

Via the megretlab organization on the Hugging Face Hub. For the companion foraging dataset, the Zenodo contact is Lizbeth Alvarado Vargas.