release/huggingface/dataset/README.md

---
license: cc-by-nc-4.0
pretty_name: DRSeg
task_categories:
  - image-segmentation
  - visual-question-answering
language:
  - en
tags:
  - uav
  - remote-sensing
  - reasoning-segmentation
  - cvpr-2026
  - pixdlm
size_categories:
  - 10K<n<100K
configs:
  - config_name: metadata
    data_files:
      - split: train
        path: metadata/train.jsonl
      - split: validation
        path: metadata/validation.jsonl
      - split: test
        path: metadata/test.jsonl
---

# DRSeg: UAV Reasoning Segmentation Benchmark

DRSeg is the dataset introduced with PixDLM for UAV reasoning segmentation. It
contains high-resolution UAV images, instance masks, reasoning questions,
reasoning answers, and reasoning-type annotations.

## Release Status

- Public release for the PixDLM CVPR 2026 Highlight work.
- The repository includes lightweight JSONL metadata for the HuggingFace Dataset
  Viewer and a full DRSeg archive for training and evaluation.
- The 2027 CVPR Compute Transparency Champion release documents split sizes,
  metrics, and evaluation assumptions in the PixDLM model repository.

## Dataset Summary

| Split | Samples | Spatial | Attribute | Scene |
| --- | ---: | ---: | ---: | ---: |
| Train | 2,999 | 999 | 1,000 | 1,000 |
| Validation | 2,000 | 667 | 667 | 666 |
| Test | 5,001 | 1,667 | 1,667 | 1,667 |

## Files

The full release is organized as:

```text
DRSeg/
├── DRtrain/
├── DRval/
├── DRtest/
└── label/
    ├── DRSeg_train.json
    ├── DRSeg_val.json
    └── DRSeg_test.json
```

The `metadata/*.jsonl` files in this repository provide lightweight preview rows
for the HuggingFace dataset viewer. They do not replace the full image/mask data.

## Task Format

Input: one UAV image and one reasoning-oriented question.

Output: the target object mask, the target category, and the answer text. The
question may require spatial, attribute, or scene-context reasoning before the
target can be identified.

## Annotation Fields

Each metadata row contains:

- `id`: sample identifier.
- `split`: train, validation, or test.
- `image`: relative image path.
- `question`: reasoning-oriented visual question.
- `answer`: CoT-style reasoning answer.
- `reasoning_type`: spatial, attribute, or scene.
- `categories`: target object categories.
- `metadata`: UAV capture metadata such as altitude, time of day, and camera angle.

## Citation

```bibtex
@inproceedings{ke2026pixdlm,
  title={PixDLM: A Dual-Path Multimodal Language Model for UAV Reasoning Segmentation},
  author={Ke, Shuyan and Mei, Yifan and Wu, Changli and Zheng, Yonghan and Ji, Jiayi and Cao, Liujuan and Ji, Rongrong},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2026}
}
```

## License

DRSeg is released for research use under CC BY-NC 4.0 unless a newer license is
specified by the authors.