Dataset Card for GZCD (Grevy's Zebra Census Dataset)

The GZCD (Grevy's Zebra Census Dataset) is an ecological computer vision dataset designed for wildlife detection, species classification, and individual animal re-identification (Re-ID). Captured entirely in the field, this dataset emphasizes endangered equid and ungulate populations, specifically Grevy's zebras, plains zebras, and reticulated giraffes.

GZCD consists of field photographs captured in Meru County, Kenya, taken over four days by 13 photographers during the 2016 and 2018 iterations of the Great Grevy's Rally (GGR). The data collection was carried out by citizen scientists and conservationists across a 25,000 square kilometer range. It was curated to help train and test new computer vision algorithms, originally as part of Jason Parham's PhD dissertation work (Parham, 2021; Parham et al., 2017) and refined by a team at the Imageomics Institute.

---

Dataset Details

This dataset provides raw ecological image data alongside the carefully processed outputs of the GGR Pipeline. It serves as an excellent resource for researchers bridging Computer Vision (CV) and ecology, providing annotations for bounding boxes, viewpoints, census annotation scores, and individual animal Re-ID clusters obtained after applying the semi-automated GGR pipeline. Manual review, while not comprehensive, has significantly increased the reliability of the results.

---

Dataset Structure

The dataset contains the raw images and two metadata files representing the data before and after processing through the GGR pipeline.

/dataset/
    images/
        <uuid 1>.jpg
        <uuid 2>.jpg
        ...
    initial_metadata.json
    final_metadata.json

Data Instances

• images/: The raw field images capturing various wildlife species.
• initial_metadata.json: Represents the original state of the dataset prior to processing, containing EXIF data, camera body details, flash card source data, GPS coordinates, and initial image dimensions.
• final_metadata.json: The fully annotated dataset resulting from the 9-stage CV pipeline, containing localized bounding boxes, species, viewpoint and census annotation classifications, and individual IDs.

Data Fields

Categories

Both JSON files share a consistent mapping of 23 categories:

category ID : category Name	category ID : category Name	category ID : category Name	category ID : category Name	category ID : category Name
0 : grevy's zebra	1 : plains zebra	2 : giraffe_reticulated	3 : ignore	4 : bird
5 : antelope	6 : impala	7 : goat	8 : water_buffalo	9 : person
10 : gazelle	11 : lion	12 : donkey	13 : domesticated_cow	14 : rhino_white
15 : elephant_savanna	16 : horse	17 : ostrich	18 : warthog	19 : car
20 : rhino_black	21 : camel	22 : ____

File: initial_metadata.json

Contains metadata of the image files before running the pipeline:

• gid: Integer global ID.
• uuid: Unique string identifier for the image. This is computed deterministically so it will not change unless the contents of the file are changed.
• uri / uri_original / original_name / original_path: File paths and names pointing to the images/ directory.
• ext: File extension (e.g., .jpg).
• width / height: Image dimensions in pixels.
• time_posix: Posix time of image capture. Cameras were time-synchronized at the start of each image collection drive.
• gps_lat / gps_lon: Geographic coordinates of the image capture. Note: Missing, masked, or unknown GPS coordinates are explicitly encoded as -1.0.
• orientation: Camera orientation integer. This represents the standard EXIF orientation tag (values 1 through 8) embedded by the camera. It tells the software which way is "up," ensuring the image isn't viewed sideways or upside-down. For example, 1 is standard/normal, 3 is upside-down, and 6 or 8 represent 90-degree rotations.
• location_code: A short categorical label identifying which data-collection campaign or site an image belongs to (e.g., "GZCD" for this dataset). It lets images from different collection events or geographic areas be grouped and filtered consistently throughout the pipeline.
• camera_make / camera_model: Hardware metadata extracted from image EXIF data.
• card_id: Identifier string tracking the specific physical storage card used during field data ingestion.

File: final_metadata.json

Contains cleaned image metadata and a rich annotations array generated by the GGR pipeline:

• Image Data: uuid, image_path, datetime, latitude, longitude (encoded as -1.0 if unknown), width, height, camera_make, camera_model, card_id.
• Annotations:
  • image_uuid: The uuid of the parent image this annotation was drawn on (i.e., the image-level uuid described in the Image Data section above). It links each detection back to its source image.
  • bbox: Bounding box of the detected animal in absolute pixel coordinates, as [x, y, width, height]. (x, y) is the top-left (upper-left) corner of the box, measured from the image's top-left origin (x increases rightward, y increases downward); width and height extend right and down from that corner.
  • region_bbox: The Census Annotation Region (CA-R) — a tighter box, in the same pixel format as bbox, that crops down to just the most identifiable part of the animal (for Grévy's zebra, the hip and shoulder chevron stripe pattern). Its purpose is to remove distracting background and body parts so that individual re-identification matches on the most discriminative region, which makes matching faster and more accurate.
  • category_id: Integer index into the species category list defined in Categories section. It maps the annotation to its species (e.g., grevy's zebra).
  • viewpoint: String giving the side of the animal shown. Possible values follow the WBIA viewpoint vocabulary: left, right, front, back, frontleft, frontright, backleft, backright. Because GZCD photographers were trained to shoot the right side, annotations are overwhelmingly right, frontright, or backright.
  • CA_score: The Census Annotation confidence score — a float from 0.0 to 1.0 output by the CA classifier, representing how confident the model is that the annotation is a clean, universally-comparable view usable for individual identification (correct side, with the identifying chevron pattern clearly visible). Higher is better: ~0.99 means a clear, high-quality view; values near 0 mean the hip or the chevron or both are snot clearly visible.
  • annotations_census: Boolean indicating whether this annotation qualifies as a Census Annotation — i.e., it passed the census filtering (correct species/viewpoint and a high enough CA_score) and is therefore considered valid for inclusion in the population count. Restricting counting to these clean views avoids over-counting from poor or ambiguous detections.
  • annot_uuid: Unique identifier for the specific detection.
  • manual: Boolean indicating whether a human reviewer set or verified this annotation. true means a person confirmed (or hand-corrected) the box and its labels; false means it was produced automatically by the detection pipeline without human review.
  • individual_id: UUID that indicates which individual animal is seen in this annotation. All annotations that were determined to be from the same individual animal will have the same unique UUID across the images..

---

Dataset Creation

Data Collection and Processing

The images were collected in Meru County, Kenya during the 2016 and 2018 census rallies. This dataset is the output of the GGR Pipeline, a 8-stage workflow designed for robust ecological tracking. Some manual corrections were made. Here are the stages:

1. Import: Parses raw field data, ingests card/camera metadata, and generates standard metadata descriptions, mapping absolute paths to the images. Synchronizations time and GPS recording with external sources.
2. Detection: Utilizes YOLO to localize animals in the images, creating initial bounding boxes (bbox).
3. Species Classification: Processes detections through BioCLIP to classify the species (e.g., Grevy's zebra vs. Plains zebra).
4. Viewpoint Classification: Determines the orientation of the detected animal (left, frontleft, back, right, backright etc).
5. Census Annotation (CA) Classification: Evaluates photographic quality and pose. Assigns a CA_score and a boolean flag indicating if the annotation shows a clear enough view of the animal's right hip and right chevron.
6. CA Filtering: Discards annotations below an identifiability threshold decided empirically and simplifies viewpoints strictly to left or right flanks for Re-ID purposes.
7. MiewID: Generates deep metric embeddings for all CA's the MiewID model (Otarashvili, 2023).
8. Local Clusters and Alternatives (LCA) Algorithm: Generates clusters of CA's each corresponding to a single animal. Uses human-in-the-loop decision making to resolve inconsistencies.

---

Considerations for Using the Data

• Missing Location Data: When GPS coordinates could not be extracted from field recordings or manual logs, fields for latitude and longitude (gps_lat/gps_lon) default to -1.0. Algorithms utilizing geometric filtering should catch these sentinel values.
• Viewpoint Filtering: Because the GGR photographers were strongly encouraged to take images showing a zebra's right hip and chevron, there is a strong bias in the dataset toward right side viewpoint.
• Geographic Focus: The imageset is collected exclusively in Meru County in Kenya. Models trained exclusively on this dataset may experience issues associated with domain shift when deployed in different environments.
• Class Imbalance: Due to the targeted focus of the Great Grevy's Rally on Grevy's Zebras, the final JSON contains individual IDs for Grevy's Zebra only.
• Focus on Identification Because the emphasis of the processing is individual animal identities, the emphasis of detection and related processing is not on completeness but instead on reliably finding and localizing.

---

Citations & References

If you use this dataset, please cite the dataset release along with the foundational computer vision work that scaled up citizen-science wildlife censusing (the GGR pipeline and Parham's dissertation) and the MiewID re-identification model used in the pipeline:

@misc{gzcd2026,
  author = {Ekaterina Nepovinnykh and Ilan Beyen and Aidan Westphal and Ankit K. Upadhyay and S M Rayeed and Julian Bain and Jason Parham and Jonathan Crall and David Kimiti and Belinda Mackey and Daniel Rubenstein and Tanya Berger-Wolf and Timm Haucke and Sara Beery and Tuomas Eerola and Heikki Kälviäinen and Charles V. Stewart},
  title = {GZCD: Grevy's Zebra Census Dataset},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/imageomics/GZCD}}
}

@phdthesis{parham2021animal,
  author = {Jason R. Parham},
  title = {Animal Detection for Photographic Censusing},
  school = {Rensselaer Polytechnic Institute},
  address = {Troy, NY, USA},
  year = {2021}
}

@inproceedings{parham2017animal,
  author = {Jason Parham and Jonathan Crall and Charles Stewart and Tanya Berger-Wolf and Daniel I. Rubenstein},
  title = {Animal Population Censusing at Scale with Citizen Science and Photographic Identification},
  booktitle = {Proceedings of the AAAI 2017 Spring Symposium on AI for Social Good},
  address = {Stanford, CA},
  year = {2017}
}

@misc{otarashvili2023miewid,
  author = {Lasha Otarashvili},
  title = {MiewID},
  year = {2023},
  doi = {10.5281/zenodo.13647526},
  url = {https://github.com/WildMeOrg/wbia-plugin-miew-id}
}

Dataset Card Authors

S M Rayeed
Elizabeth G. Campolongo