## Motivation

Being able to tell **which individual animal is which** supports a lot of real wildlife work—estimating population size, tracking movement, and studying behavior—because it links photos back to the *same* animal over time. But as camera traps, drones, and citizen-science apps collect more images, manual matching doesn’t scale. The data grows much faster than the number of experts who can label it.

Many benchmarks focus on **matching**: given a photo, decide whether it belongs to one of the individuals already stored in a catalog (closed-set or open-set matching). In the field, that’s only part of the job. You also need **discovery**: when the catalog is incomplete (which is common), you must recognize when a photo shows a **new individual** and then group those new photos so each group corresponds to one animal that can be added to the database.

That’s what this competition targets. You will work in a realistic, multi-species setting where the test set includes both:
- **known individuals** (for species with a provided identity database), and
- **new individuals** that must be **discovered and grouped**.

This is challenging because the same animal can look very different across images (pose, lighting, viewpoint, aging, injuries), backgrounds can be cluttered, cameras and locations vary, and some individuals have many more images than others. The goal is to encourage methods that can help build and update animal ID databases with minimal manual work.

## What will you learn?

You’ll learn the basics of **animal re-identification with discovery**—how to represent individuals, handle dataset shifts, and build a pipeline that can both **match** images to known identities and **group** images of new individuals. You’ll also learn how the results are scored using the **Adjusted Rand Index (ARI)**, which rewards correct grouping and penalizes both splitting one individual into multiple groups and merging different individuals into the same group.