# DRSeg Data

DRSeg is a UAV reasoning segmentation benchmark with 10,000 image-level samples.
Each sample contains a UAV image, one or more segmentation annotations, a
reasoning question, a reasoning answer, and a reasoning type.

## Splits

| Split | Samples |
| --- | ---: |
| Train | 2,999 |
| Validation | 2,000 |
| Test | 5,001 |

## Expected Layout

```text
data/DRSeg/
├── DRtrain/
├── DRval/
├── DRtest/
├── label/
│   ├── DRSeg_train.json
│   ├── DRSeg_val.json
│   └── DRSeg_test.json
├── CODrone -> .
└── labels -> label
```

The `CODrone` and `labels` entries are compatibility links for the original
dataset loader.

## Reasoning Types

- `spatial`: position and spatial relation reasoning.
- `attribute`: visual attribute and object property reasoning.
- `scene`: scene-context reasoning.

## Metadata Preview

The HuggingFace dataset repo includes lightweight metadata JSONL files under
`metadata/`. They are intended for dataset-card preview and quick inspection.
Use the full image/mask archive for training and evaluation.