# Cup-ybara — Data

来源：https://www.kaggle.com/competitions/cupybara/data

本地 Meta Kaggle 快照；不是本次实时页面。

# Dataset Description

# Overview

The dataset for this competition is composed of real-world footage collected by camera traps in Uruguayan ecosystems. It includes:

- Unlabeled training data (train/)
- Labeled test data for public leaderboard evaluation (test/)
- Held-out private test data for final scoring (private_test/)

Each video is exactly 15 seconds long, encoded in .mp4 format, and recorded using motion-triggered camera traps deployed in natural, uncontrolled environments.

**File Structure**

```text
├── train/
│   ├── 000123.mp4
│   ├── 000124.mp4
│   └── ...
│
├── test/
│   ├── 100234.mp4
│   ├── 100235.mp4
│   └── ...
│
├── test.csv         # Species labels for all test videos
```

---

# Training Data

Folder: `train/`

- Contains approximately 2,000 unlabeled videos.
- Videos are drawn from multiple environments, times of day, and camera devices.
- Participants must generate their own labels (e.g., via pseudo-labeling or self-supervised learning).
- There is no metadata or annotations provided for these videos.

> ⚠️ Important: The distribution of classes, lighting, movement, and background in the train set does not necessarily match the distribution in the test or private datasets.

---

# Test Data

Folder: `test/`
Labels: `test.csv`

- Contains 412 labeled videos.
- Used to generate scores for the public leaderboard during the competition.
- Ground truth labels are provided via test.csv, which includes the full labels (including those for the private leaderboard). The public/private split is not disclosed.
- Submissions are compared against the appropriate portion of this set, but participants cannot distinguish between public and private examples during development.

> 📌 All filenames in your submission must match those in test.csv, including casing and extension.

---

# Private Test Set

Name: `private_test/`

- Contains an additional set of labeled videos not released during the competition.
- After the competition closes, all models that achieved a public leaderboard score ≥ 0.40 will be re-evaluated offline on this dataset.
- This score determines the final ranking and prize eligibility.

>🔐 The videos and labels in private_test will only be shared publicly after the competition ends.

---

# Video Format

- Format: .mp4
- Duration: 15 seconds
- Resolution and framerate may vary slightly depending on the camera used
- Some videos may appear to contain no animals or show only subtle movement (e.g., insects or distant animals)
- Lighting conditions vary (day/night, sun/shade)

---

# Class Labels

Participants must classify each video into one of the following 14 classes:

```bash
'armadillo'
'bird'
'capybara'
'cow'
'dusky_legged_guan'
'gray_brocket'
'hare'
'human'
'insect'
'margay'
'no_animal'
'skunk'
'unknown_animal'
'wild_boar'
```

🔸 Use single quotes (') around each class label in your submission.
🔸 The 'no_animal' class includes videos where no animal is visible or identifiable.
🔸 The 'unknown_animal' class includes videos where an any other animal is present, but is not from one of the categories above.

---

# Best practices and Notes

- The class distribution is unbalanced, and some species are significantly underrepresented.
- Backgrounds, camera angles, and motion patterns can vary widely.
- Some classes (e.g., 'insect', 'no_animal', 'unknown_animal') may be visually ambiguous.
- Participants are encouraged to explore transfer learning, temporal modeling, and weak supervision to maximize performance.
- Use the provided demo notebook and model interface as a starting point to ensure submission compatibility.