---
license: apache-2.0
---

# **NICO++ DG Benchmark Subset (Unofficial)**

## Dataset Summary

This is a **non-official subset** of the [NICO++ dataset](https://arxiv.org/abs/2204.08040), designed for **Domain Generalization (DG)** evaluation.
We  select **20 categories** across **6 domains**.

The dataset can be used to benchmark algorithms for **domain generalization, domain adaptation, and robustness testing**.

⚠️ **Note:** This dataset is **not the official release of NICO++**, but a re-organized subset curated for research purposes.

---

## Supported Tasks and Leaderboards

* **Domain Generalization (DG)**
* **Out-of-Distribution (OOD) Robustness**
* **Representation Learning with Multiple Contexts**

---

## Languages

* Images contain natural objects and scenes; no text annotations.
* Labels are in **English**.

---

## Dataset Structure

### Data Fields

Each sample contains:

* `image`: the input image (RGB)
* `label`: the class label (integer)
* `category`: the semantic category (string, one of 20)
* `domain`: the environment/domain (string, one of 6)

### Domains

We follow the DG benchmark setup:

```python
"domains" = ['autumn', 'dim', 'grass', 'outdoor', 'rock', 'water']
```

### Categories

The selected 20 categories are:

```python
"categories" = [
 'kangaroo', 'dolphin', 'sailboat', 'pumpkin','gun','sheep','tent','mailbox','cactus','car',
 'spider','tortoise','fox','lion','elephant','racket','umbrella','crab','giraffe','chair'
]
```

---

## Data Splits

The dataset is split into domains rather than standard train/val/test.

* Researchers may adopt **leave-one-domain-out** DG evaluation, where training uses 5 domains and testing uses the held-out one.
* Example: Train on {autumn, dim, grass, outdoor, rock}, Test on {water}.