---
title: NJU-HARD-Detection
canonical_url: "https://www.modelscope.cn/datasets/KAIWANG/NJU-HARD-Detection"
md_url: "https://www.modelscope.cn/datasets/KAIWANG/NJU-HARD-Detection.md"
repository: KAIWANG/NJU-HARD-Detection
chinese_name: NJU-HARD-Detection
last_updated: 2026-09-18
license: unknown
downloads: 10452
stars: 0
---

# NJU-HARD-Detection

> NJU-HARD-Detection - KAIWANG 在 ModelScope 开源的数据集。HARD detection task release: 3,549 original 122 MP UAV frames, 21 sequences, 138,830 box annotations; native Parquet with manuscript sequence splits and preserved source labels. HARD 目标检测任务版：3,549 张原始 1.22…

KAIWANG/NJU-HARD-Detection 是 ModelScope 魔搭社区上的数据集，采用 unknown 许可。

- **Repository**: KAIWANG/NJU-HARD-Detection
- **License**: unknown
- **Downloads**: 10452
- **Stars**: 0
- **Last updated**: 2026-09-18

Source: https://www.modelscope.cn/datasets/KAIWANG/NJU-HARD-Detection

---

# NJU-HARD-Detection

Original aerial images and object annotations for detection research.

The complete release contains 3,549 frames from 21 sequences and 138,830 box annotations. The four categories are pedestrian, car, bus, and truck. Original images are 12,768 by 9,564 pixels.

The `default` configuration provides the original-resolution train, validation, and test splits. The `preview` configuration contains seven smaller sample images.

Each Parquet row contains an `image`, frame and sequence identifiers, and an `objects` list. Boxes use pixel coordinates in `[x, y, width, height]` format. Source annotations are preserved. Five frames have missing labels (`objects=null`); source ID and box issues are recorded in `quality_flags`.

```python
from modelscope.msdatasets import MsDataset

data = MsDataset.load(
    "KAIWANG/NJU-HARD-Detection",
    subset_name="default",
    split="train",
    use_streaming=True,
)
sample = next(iter(data))
```

The source documentation specifies research use. Complete license terms were not supplied, so the license metadata is `unknown`.

[中文说明](https://www.modelscope.cn/datasets/KAIWANG/NJU-HARD-Detection/file/view/master/README_ZH.md?status=1)