Gorilla-SPAC-Wild: Large-Scale Video Dataset for Gorilla Re-Identification

Overview

Gorilla-SPAC-Wild is a comprehensive benchmark dataset for individual re-identification of Western Lowland Gorillas from camera trap footage in natural rainforest environments. This dataset addresses a critical bottleneck in conservation: automating the analysis of vast archives of camera trap video to track endangered gorilla populations non-invasively.

This dataset is part of the GorillaWatch project, which introduces an end-to-end pipeline integrating detection, tracking, and re-identification for automated gorilla monitoring.

Dataset Description

Key Features

• Large-Scale Video Dataset: Extracted from camera trap footage at Odzala-Kokoua National Park, Republic of Congo
• 135 Individual Gorillas: Tracked across multiple encounters in natural, challenging lighting and environmental conditions
• Paired Face-Body Crops: Each gorilla is annotated with both facial crops and full-body crops, enabling comprehensive re-identification analysis
• Cross-Encounter Data: Gorillas recorded in different locations and at different times, representing real-world re-identification challenges where appearance changes due to varying environmental conditions
• Open-Set Evaluation: Strict individual-based split suitable for real-world scenarios where new individuals frequently appear

Example Images

These images show some examples of the same dataset entry in the different dataset configs. The large one is the full frame (content of the image column in the two full_image configs) with the corresponding bounding boxes for face and body drawn onto it. Next to it are the face and body crops, which can be found in the image and/or body_image columns in the body and face_with_body configs.

Data Extraction & Processing

• Images are extracted from automatically generated tracklets using a YOLOv8-Nano model, which has been finetuned for gorilla detection
• Ground truth labels provided by primate researchers with 15+ years of field experience with this population
• For cropping, the bounding boxes were squared to avoid distortion of important facial features in the preprocessing's resizing

Dataset Splits

• Train (70%): For training re-identification models
• Validation (15%): For hyperparameter tuning and model selection
• Test (15%): For final evaluation on unseen encounters
• Single-Encounter: All gorillas with only a single encounter, thus not properly usable as cross-encounter data

The splits contain disjoint classes, to evaluate on unseen gorillas and allow for better real-world comparability.

Use Cases

1. Individual Re-Identification: Training models to identify individual gorillas from camera trap footage
2. Conservation Monitoring: Automating population tracking for demographic studies
3. Disease Tracking: Monitoring health and injury patterns in wildlife populations
4. Behavioral Analysis: Long-term observation of social dynamics

Dataset Schema

Across all configs, each sample includes:

• image: Primary image (depending on config, encoded as image bytes)
• class: Individual gorilla identifier
• date: Capture date (YYYYMMDD format)
• video: Source video identifier
• frame_number: Frame number within the video
• camera: Camera trap identifier

Depending on the config, there are slight differences in the available columns and their contents:

• body:
  • image: Body crop
• face_with_body:
  • image: Face crop
  • body_image: Corresponding body crop. (The intention behind this is, to have subset of the body dataset, where faces are visible.)
• full_image_bbox_body:
  • image: Full video frame
  • bbox: Detected body bounding box (in the format [x_min,y_min, width, height])
• full_image_bbox_face_with_body:
  • image: Full video frame
  • bbox: Detected face bounding box
  • body_bbox: Detected body bounding box

Note that the full_image configs do not contain all detection entries for a frame in a single entry. If there are multiple gorillas detected and annotated in the same frame, there will be multiple dataset entries for that frame, each containing one gorilla. There are also frames where not all visible gorillas are annotated, due to faulty detection and tracking or uncertainties in identifying them.

Dataset Statistics

• Total Samples: 160,818 annotated gorilla face/body image pairs and 353,820 annotated gorilla body images

  Split	Percentage	Classes	Face Images	Body Images
  Training	~70%	75	110,411	242,420
  Validation	~15%	17	23,021	49,901
  Test	~15%	16	22,619	52,014

  
• Individual Count: 135 different gorillas, 108 in the train, test and validation data. However, there is quite some imbalance between the classes' sample counts:
• Temporal Span: Multiple years of camera trap recordings
• Locations: 33 different camera locations in the Odzala-Kokoua National Park, Republic of Congo

Source Videos

Additionally to the prepared dataset, we provide the source videos from which the images are taken. They are provided in the videos.tar.gz archive in this repo and can be unpacked with tar -xzf videos.tar.gz.

Loading and Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("gorilla-watch/Gorilla-SPAC-Wild", "face_with_body")

# Split structure
train_data = dataset["train"]
val_data = dataset["validation"]
test_data = dataset["test"]

# Access samples
sample = train_data[0]
face_image = sample["image"]  # PIL Image object
gorilla_id = sample["class"]   # Individual identifier

Ethical Considerations

• This dataset is used exclusively for non-invasive wildlife conservation and monitoring
• All data collection followed ethical guidelines for field research
• The dataset supports critical conservation efforts for an endangered species

License

This dataset is released under the CC-BY-4.0 License.

Citation

If you use this dataset in your research, please cite:

@inproceedings{GorillaWatch2026,
  title={GorillaWatch: An Automated System for In-the-Wild Gorilla Re-Identification and Population Monitoring}, 
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  author={Maximilian Schall and Felix Leonard Knöfel and Noah Elias König and Jan Jonas Kubeler and Maximilian von Klinski and Joan Wilhelm Linnemann and Xiaoshi Liu and Iven Jelle Schlegelmilch and Ole Woyciniuk and Alexandra Schild and Dante Wasmuht and Magdalena Bermejo Espinet and German Illera Basas and Gerard de Melo},
  year={2026},
  archivePrefix={arXiv},
  eprint={2512.07776}
}

Acknowledgments

The project on which this report is based was funded by the Federal Ministry of Research, Technology and Space under the funding code “KI-Servicezentrum Berlin-Brandenburg” 16IS22092. We acknowledge the support of Sabine Plattner African Charities (SPAC) for their funding to this research. We are grateful to Zoo Berlin for their expert assistance and facility access. This collaboration enabled the development of AI tools capable of being deployed in the wild to directly support gorilla conservation. The responsibility for the content of this publication remains with the authors.