red_bee_reID: Re-Identification of Red Painted Honey Bees

A curated re-identification dataset of 45 identities of honey bees (Apis mellifera) painted with red paint on the thorax. On average, each identity has 6.2 tracklets for a total of 358 tracks representing 9495 image crops, and an average of 24.1 images/track. Both raw, unrotated images and standardized background mask crops are provided.

Dataset Details

This ReID dataset was generated from 9 videos captured over 3 days at the entrance ramp of a bee feeder at various locations in the north of Puerto Rico. The raw videos (not included) show the bees entering from the bottom, walking, and exiting from the top of the screen. Video was obtained using a Basler camera (model a2A3840-45ucPRO) with resolution 3840x2160 px at 17 fps and encoded as H265 videos. As many bees as possible were marked with one dot of paint on the thorax (red or green). Nectar foragers typically come back through the entrance several times over the span of multiple days, thus providing realistic multi-day ReID data.

Using the methods explained in section 4.1 of the article , all bees were detected, tracked, and their crops extracted. Crops were extracted at 1.5 the pixel length of the average bee skeleton per video, so images may vary slightly in size. Crops were rotated so the angle between the head and abdomen keypoints were vertical. Then the identity of painted bees was manually annotated. For the purpose of the study, we kept only the identities with 3 or more tracks, and were painted with red color, leaving 45 identities. On average, each identity has 6.2 tracks for a total of 358 tracks representing 9495 image crops, and an average of 24.1 images/track. Train test splits using this dataset should not split within tracks to prevent data leakage.

Supported Tasks and Leaderboards

This data was published according with the paper One-Shot Fine-Grained Re-Identification of Paint Marked Honey Bees using Vision Foundation Models for one shot reidentification using deep metric learning. Published method achieves ~85% top1 accuracy contrastively trained using a single track of training in the closed set.

Dataset Structure

Images are organized by pretreatment method, inside an images folder. Parquet format has been used to reduce size on remote storage repo.

/images
    rotated_masked  
        bf-rg_2025-04-08_08.cfr.mp4.T000189_F005007.jpg
        bf-rg_2025-04-08_08.cfr.mp4.T000189_F005008.jpg
        ...
    unrotated_unmasked
        bf-rg_2025-04-08_08.cfr.mp4.T000189_F005007.jpg
        bf-rg_2025-04-08_08.cfr.mp4.T000189_F005008.jpg
        ...
metadata.csv

Data Instances

Image filenames are named following their data collection site, date, tracking id, and frame from video detection. For instance: bf-rg_2025-04-08_08.cfr.mp4.T000189_F005007.jpg:

• bf : bee_feeder (experimental setup shortname)
• rg : Rio Grande (site shortname)
• 2025-04-08 : date of collection
• 08.cfr : video number + cfr(constant frame rate)
• .mp4 : video format
• T000189 : track 189 (track ids are new generated per video)
• F005007 : detection on frame 5007
• .jpg : image format

Video detection and tracking are described in section 4.1

Rotated, maksed images were extracted at 1.5 the pixel length of the average bee skeleton per video, so images may vary slightly in size. Subsequently, images were rotated so the angle between the head and abdomen keypoints were vertical, and background masking was performed using SAM2, with input points from detected skeleton. Unrotated, unmasked images were re-extracted from video at the same size as their preprocessed counterpart, but simple centered on the waist keypoint.

Data Fields

metadata.csv:

• rotated_masked : path to rotated and masked (preprocessed) crop image.
• unrotated_unmasked : path to original unrotated, unmasked crop image.
• frame : frame number within the source video.
• det_id : detection ID within the frame.
• score : detection confidence score.
• video_name : name of the source video file.
• bbox_x : x-coordinate of bounding box (top-left corner).
• bbox_y : y-coordinate of bounding box (top-left corner).
• bbox_width : width of bounding box.
• bbox_height : height of bounding box.
• bbox_confidence : confidence score of bounding box detection.
• head_x : x-coordinate of head keypoint.
• head_y : y-coordinate of head keypoint.
• head_conf : confidence score of head keypoint.
• head_visibility : visibility flag for head keypoint.
• neck_x : x-coordinate of neck keypoint.
• neck_y : y-coordinate of neck keypoint.
• neck_conf : confidence score of neck keypoint.
• neck_visibility : visibility flag for neck keypoint.
• thorax_x : x-coordinate of thorax keypoint.
• thorax_y : y-coordinate of thorax keypoint.
• thorax_conf : confidence score of thorax keypoint.
• thorax_visibility : visibility flag for thorax keypoint.
• waist_x : x-coordinate of waist keypoint.
• waist_y : y-coordinate of waist keypoint.
• waist_conf : confidence score of waist keypoint.
• waist_visibility : visibility flag for waist keypoint.
• tail_x : x-coordinate of tail keypoint.
• tail_y : y-coordinate of tail keypoint.
• tail_conf : confidence score of tail keypoint.
• tail_visibility : visibility flag for tail keypoint.
• waist_src : source method used to estimate waist position.
• angle_src : source method used to estimate orientation angle.
• track_id : tracker-assigned ID within video.
• cx : x-coordinate of object center.
• cy : y-coordinate of object center.
• angle : orientation angle of the object.
• virtual : whether detection is interpolated (virtual).
• vx : estimated velocity in x-direction.
• vy : estimated velocity in y-direction.
• cost : tracking association cost value.
• pcx : predicted x-coordinate of center.
• pcy : predicted y-coordinate of center.
• pa : predicted orientation angle.
• pvx : predicted velocity in x-direction.
• pvy : predicted velocity in y-direction.
• dorsal_len : estimated dorsal (body) length.
• crop_bee_scale : scale factor used for crop resizing.
• crop_filename : filename of saved crop image.
• global_track : globally unique track ID across dataset.
• unique_track_hash : unique hash identifier for track.
• global_frame : globally indexed frame number.
• unique_frame_hash : unique hash identifier for frame.
• date : recording date of the video.
• vid_num : video number identifier.
• track_len : total length of the track (in frames).
• local_track : track ID within source video.
• real_crop_filename : filename of original (non-augmented) crop.
• new_filepath : updated file path for processed data.
• key : unique row identifier.
• bee_id : assigned bee identity label.
• video_key : unique identifier for video.
• is_track_ref : whether sample is a reference track.
• is_bee_id_ref : whether sample is a reference bee ID.

Data Splits

Train/test splits were made at the track level to prevent data leakage. 10 trials were generated by following a one-shot approach where for each ID, one track is sampled at random as training/reference, all the other tracks being used as test.

Dataset Creation

Data was collected across various days at Escuela Superior Pedro Falu Orellano in Rio Grande, PR. Bees were trained to come to "beefeeder" setups as part of the AC3 Bee Hunting project and bees were painted with help of student volunteers. Bees passing into the feeder were recorded on video, and paint was applied at nearby auxillary sugar solution feeders. The article details the subsequent processing pipeline for video data.

Curation Rationale

Individual identification of honeybees is necessary in order to study in detail the behavior of these critical pollinators. In field experiments paint codes of one or two colors that can be read by researchers are useful, and are readily identifiable by computer vision systems given sufficient training data. Single color paint marks, that rely on random variations in paint shape and placement, are a solution to the limited verbosity of structured approaches and an intermediate step towards markerless biometric ReID.

Annotations

Identity labels were annotated at the track level using a custom annotation tool. A set of reference images per track was carefully accumulated and used to match incoming tracks. An estimated 6-10 hours of annotation effort was necessary to reach 45 identities that passed the minimum threshold of 3 or more tracks.

Personal and Sensitive Information

The authors have no knowledge of sensitive information contained in the currently published dataset.

Considerations for Using the Data

Bias, Risks, and Limitations

While no tracking errors were detected while working with the data, they may occur. SAM2 background segmentation was run completely automatedly, and was not verified. As such, background masking may be slightly noisy.

As identity was annotated at track level, not all images necessarily contain sufficient information to perfrom re-identification. Indeed, some images are known to be explicitly not identifiable, i.e. bee is upside down. Virtual detections have been parsed from publicly present data.

-->

Licensing Information

This dataset is dedicated to the public domain for the benefit of scientific pursuits. We ask that you cite the dataset and journal paper using the below citations if you make use of it in your research.

Citation

Paper

@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}
}

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, and special thanks is extended to L. Alvarado Vargas, A. Rodriguez and M. Geria. This work used the UPR High-Performance Computing facility, supported by NIH/NIGMS, award 5P20GM103475

Dataset Card Authors

Luke Meyers

Dataset Card Contact

luke.meyers@upr.edu