Dataset Card for MetaWild

The MetaWild dataset is a multimodal benchmark designed to improve animal re-identification (Animal ReID) by integrating environmental metadata with visual data. Unlike existing datasets that rely solely on images, MetaWild includes 20,890 images spanning six representative species, each image is paired with metadata such as temperature and circadian rhythms, providing valuable context for distinguishing individual animals.

Dataset Details

MetaWild contains 20,890 images spanning 6 representative animal species commonly seen in the New Zealand wilderness:

• Deer
• Hare
• Penguin
• Stoat
• Wallaby
• Pūkeko

Each image is paired with both:

• Identity label (individual animal ID)
• Environmental metadata, including:
  • face_direction: one of {0: front, 1: back, 2: left, 3: right}
  • temperature: numeric value in Celsius
  • circadian rhythms: binary label indicating circadian context (day/night)
  • timestamp: time of image capture All images were captured by trail cameras deployed across diverse ecological zones, and are stored per species as structured .json files.

Species Selection

The six species in MetaWild were carefully selected to represent different ecological roles and conservation priorities in New Zealand:

• Stoat: Invasive predator threatening native wildlife
• Yellow-eyed Penguin: Endangered native species requiring urgent protection
• Wallaby & Hare: Agricultural pests causing ecosystem damage
• Deer & Pūkeko: Native species with diverse behavioral patterns

This selection ensures broad ecological coverage and practical relevance for wildlife monitoring applications.

---

Dataset Construction

To ensure high-quality, metadata-rich animal ReID data, we applied the following principles:

• Ecological Diversity: We selected six species balancing invasive/native status and appearance diversity.
• Image Quality: Images with blurred, low-light, or occluded views were filtered out.
• Metadata Availability: Each image was required to contain embedded metadata (temperature, time, etc.) extracted from camera overlays or context inference.
• Minimum Representation: Each species has at least 2,000 images and dozens of unique identities to enable identity-level modeling.

The dataset was derived from the publicly available NZ-TrailCams repository.

---

Licensing & Access

The MetaWild dataset inherits its licensing terms from the NZ-TrailCams dataset, from which it was constructed.
Specifically, MetaWild complies with the:

Community Data License Agreement – Permissive – Version 1.0
Full License Text (CDLA-Permissive-1.0)

To respect the original licensing conditions and minimize storage duplication, we do not host or redistribute any original images from the NZ-TrailCams repository.
Instead, MetaWild only includes structured .json files that contain:

• Cleaned and annotated metadata for each image
• Image filenames in the format:
  $speciesID$-$cameraID$-$imageCount$
• Identity labels and environmental metadata (e.g., temperature, circadian time, face orientation)

If access to the original images is required, users are encouraged to:

• Visit the official NZ-TrailCams repository
• Use the image filenames provided in our JSON files to retrieve specific samples by matching:
  • Species Name
  • Camera ID
  • Frame Count / Index

This approach ensures reproducibility while respecting original data ownership and storage constraints.

---

Metadata Distribution Visualizations

Temperature

Circadian Rhythms

Face Orientation

Resources

• MetaWild Website
• MetaWild Code