mpg-ranch/wolverines

Wolverine re-identification from camera trap imagery

Images and labels for a manuscript under review at Ecological Informatics on automated re-identification and novelty detection of wolverines (Gulo gulo) from bait-station camera trap images. Code, runbooks, and results are in the GitHub repository mosscoder/mpg-wolverines.

Project Overview

Wolverines are listed as threatened under the U.S. Endangered Species Act and occur at low density in remote terrain, so camera trap re-identification is one of few practical ways to study them. A bait frame exposes the ventral pelage pattern to a nearby game camera, but the animal moves while it feeds and the pattern is clearly visible in only a fraction of images. The pipeline has two stages.

1. Pelage visibility classifier. An image classifier scores every image from 0 to 1, the estimated probability that the pelage pattern is clearly visible. We call this the quality score.
2. Re-identification with novelty detection. A frozen image encoder turns each image into an encoder output, a trained projection head maps that output to an embedding, and images are compared by the cosine similarity of their embeddings. A query is assigned to the known individual of its nearest gallery image, or flagged as unknown when that similarity falls below a threshold. We compared three encoders: DINOv3-ViT-B/16 (general-purpose; Siméoni et al. 2025), BioCLIP-2 ViT-L/14 (biology-specific; Gu et al. 2025), and MegaDescriptor-L-384 (wildlife-specific; Čermák et al. 2023).

Quality score thresholds of 0, 0.25, and 0.50 are applied to the gallery, to the queries, or to both. Re-identification is scored by recall at rank one and novelty detection by balanced accuracy, each averaged over individuals. There are two experiments, a few-shot experiment (2 to 64 gallery images per individual, eight seeds) and a full-data experiment that trains on every eligible gallery image. Both report performance on the test split, the most recent events of each known individual, at checkpoints selected on the validation queries.

The quality score as a gradient within a single camera trap event. Each row is one daytime event of one individual, ten images sorted from low to high score, with the score printed on each image. Camera, scene, and lighting are fixed within a row, so the score changes with the animal's pose alone.

Configurations

Every image is a MegaDetector crop of one wolverine. Splits are temporal within each individual, so no test image predates a training image of the same animal. The dataset carries no coordinates.

pelage holds the human-labeled crops used to train the pelage visibility classifier, 2,431 training and 270 test images from the earliest camera trap events of each individual. label is the annotator's three-level rating of the ventral pelage pattern (0 None, 1 Partial, 2 Full), and the classifier treats Full as visible and the other two as not visible. Other columns give the individual (id), the event start as YYYYMMDDHHMM (ymdh), color or black-and-white infrared capture (color, 1 or 0), the source filename, and the MegaDetector confidence and bounding box.

reidentification holds every scored crop from the remaining events, 44,201 training and 5,111 test images from 11 individuals, with the classifier's pelage_score (the quality score) and the individual's identity. The test split is the most recent 10% of each individual's events. Which individuals serve as known individuals, which as unknown individuals, and which events become validation queries is decided downstream, as documented in the repository's role assignment runbook.

from datasets import load_dataset
ds = load_dataset("mpg-ranch/wolverines", "reidentification", split="train")

How the dataset was built from raw camera trap images is documented in the repository's dataset creation runbook.

Runbooks

Each pipeline stage has a README in the GitHub repository that lists its scripts in run order, their inputs and outputs, and the values used for the manuscript.

1. Dataset creation
2. Pelage visibility classifier
3. Role assignment and manuscript assets
4. Re-identification and novelty detection

Citation

Cite the dataset as:

Doherty, K., Baughan, K., Davis, B., and Ramsey, P. 2026. wolverines. Hugging Face. https://doi.org/10.57967/hf/10575

@misc{kyle_doherty_2026,
    author    = { Kyle Doherty and Kalon Baughan and Bret Davis and Philip Ramsey },
    title     = { wolverines (Revision 5316a09) },
    year      = 2026,
    url       = { https://huggingface.co/datasets/mpg-ranch/wolverines },
    doi       = { 10.57967/hf/10575 },
    publisher = { Hugging Face }
}

The manuscript citation will be added on publication.

References

• Čermák, V., Picek, L., Adam, L., and Papafitsoros, K. 2023. WildlifeDatasets: an open-source toolkit for animal re-identification. arXiv:2311.09118. https://doi.org/10.48550/arXiv.2311.09118
• Gu, J., Stevens, S., Campolongo, E. G., Thompson, M. J., Zhang, N., Wu, J., Kopanev, A., Mai, Z., White, A. E., Balhoff, J., Dahdul, W., Rubenstein, D., Lapp, H., Berger-Wolf, T., Chao, W.-L., and Su, Y. 2025. BioCLIP 2: emergent properties from scaling hierarchical contrastive learning. arXiv:2505.23883. https://doi.org/10.48550/arXiv.2505.23883
• Hernandez, A., Miao, Z., Vargas, L., Beery, S., Dodhia, R., Arbelaez, P., and Lavista Ferres, J. M. 2024. Pytorch-Wildlife: a collaborative deep learning framework for conservation. arXiv:2405.12930. https://doi.org/10.48550/arXiv.2405.12930
• Siméoni, O., Vo, H. V., Seitzer, M., Baldassarre, F., Oquab, M., Jose, C., Khalidov, V., Szafraniec, M., Yi, S., Ramamonjisoa, M., Massa, F., Haziza, D., Wehrstedt, L., Wang, J., Darcet, T., Moutakanni, T., Sentana, L., Roberts, C., Vedaldi, A., Tolan, J., Brandt, J., Couprie, C., Mairal, J., Jégou, H., Labatut, P., and Bojanowski, P. 2025. DINOv3. arXiv:2508.10104. https://doi.org/10.48550/arXiv.2508.10104