# ADE20K Sky-Water-Person Split Files

These split files list ADE20K images that contain at least one of the
target classes (sky, water, person).

## Format

Each line is a path relative to the ADE20K image root
(`ADEChallengeData2016/images/`):

```
validation/ADE_val_00000001.jpg
training/ADE_train_00000001.jpg
```

## Expected Dataset Layout

```
ADEChallengeData2016/
├── images/
│   ├── training/
│   │   ├── ADE_train_00000001.jpg
│   │   └── ...
│   └── validation/
│       ├── ADE_val_00000001.jpg
│       └── ...
└── annotations/
    ├── training/
    │   ├── ADE_train_00000001.png
    │   └── ...
    └── validation/
        ├── ADE_val_00000001.png
        └── ...
```

## Class Mapping

ADE20K raw labels → skywater 4-class scheme:

| ADE20K ID | ADE20K Name | Target Class |
|-----------|-------------|--------------|
| 3 | sky | 1 (Sky) |
| 13 | person | 3 (Person) |
| 22 | water | 2 (Water) |
| 27 | sea | 2 (Water) |
| 61 | lake | 2 (Water) |
| 105 | pond | 2 (Water) |
| 110 | fountain | 2 (Water) |
| 114 | waterfall | 2 (Water) |
| 129 | river | 2 (Water) |
| (all others) | — | 0 (Background) |

## Obtaining ADE20K

Download from: https://groups.csail.mit.edu/vision/datasets/ADE20K/

Use the `ADEChallengeData2016` subset (training + validation).