README.md

---
pretty_name: STAR Relationship
tags:
- scene-graph-generation
- remote-sensing
- polygon-annotation
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
dataset_info:
  features:
  - name: id
    dtype: string
  - name: image
    dtype: string
  - name: width
    dtype: int32
  - name: height
    dtype: int32
  - name: polygons
    list:
      list:
        list: float32
        length: 2
  - name: labels
    list:
      class_label:
        names:
          '0': __background__
          '1': airplane
          '2': apron
          '3': arch_dam
          '4': baseball_diamond
          '5': basketball_court
          '6': boarding_bridge
          '7': boat
          '8': breakwater
          '9': bridge
          '10': car
          '11': car_parking
          '12': cement_concrete_pavement
          '13': chimney
          '14': coal_yard
          '15': containment_vessel
          '16': cooling_tower
          '17': crane
          '18': dock
          '19': engineering_vehicle
          '20': flood_dam
          '21': foundation_pit
          '22': gas_station
          '23': genset
          '24': goods_yard
          '25': gravity_dam
          '26': ground_track_field
          '27': intersection
          '28': lattice_tower
          '29': roundabout
          '30': runway
          '31': ship
          '32': ship_lock
          '33': smoke
          '34': soccer_ball_field
          '35': stadium
          '36': storehouse
          '37': substation
          '38': tank
          '39': taxiway
          '40': tennis_court
          '41': terminal
          '42': toll_gate
          '43': tower_crane
          '44': truck
          '45': truck_parking
          '46': unfinished_building
          '47': vapor
          '48': wind_mill
  - name: relations
    struct:
    - name: subject_index
      list: int64
    - name: object_index
      list: int64
    - name: predicate
      list:
        class_label:
          names:
            '0': __background__
            '1': adjacent
            '2': approach
            '3': around
            '4': away from
            '5': co-storage with
            '6': connect
            '7': converge
            '8': directly connected to
            '9': directly transmit electricity to
            '10': docked alongside with
            '11': docking at the different dock with
            '12': docking at the same breakwater with
            '13': docking at the same dock with
            '14': drive off
            '15': drive toward
            '16': driving alongside with
            '17': driving in the different lane with
            '18': driving in the opposite direction with
            '19': driving in the same direction with
            '20': driving in the same lane with
            '21': exhaust to
            '22': in the different parking with
            '23': in the same parking with
            '24': incorrectly parked on
            '25': indirectly connected to
            '26': indirectly transmit electricity to
            '27': intersect
            '28': isolatedly docked at
            '29': isolatedly parked on
            '30': not co-storage with
            '31': not docked alongside with
            '32': not parked alongside with
            '33': not run along
            '34': not working on
            '35': over
            '36': parallelly docked at
            '37': parallelly parked on
            '38': parked alongside with
            '39': parking in the different apron with
            '40': parking in the same apron with
            '41': pass across
            '42': pass through
            '43': pass under
            '44': randomly docked at
            '45': randomly parked on
            '46': run along
            '47': running along the different runway with
            '48': running along the different taxiway with
            '49': running along the same taxiway with
            '50': slightly emit
            '51': supply to
            '52': through
            '53': violently emit
            '54': within danger distance of
            '55': within different line of
            '56': within safe distance of
            '57': within same line of
            '58': working on
  splits:
  - name: train
    num_bytes: 10717079
    num_examples: 771
  - name: validation
    num_bytes: 4288100
    num_examples: 238
  - name: test
    num_bytes: 15840
    num_examples: 264
  download_size: 14671494
  dataset_size: 15021019
---

# STAR Relationship

STAR Relationship 是一个遥感场景图生成（Scene Graph Generation，SGG）数据集。仓库将原始大尺寸图片与结构化标注分开保存：

- 图片以普通文件形式位于 `images/`。
- train、validation 和 test 的结构化标注以 Parquet 保存。
- Dataset 中的 `image` 字段是图片相对于仓库根目录的路径，不包含图片字节，也不会自动解码为 PIL 对象。

完整仓库约为 127 GB。使用 `snapshot_download()` 下载完整仓库前，请确认本地有足够的磁盘空间。

## 仓库结构

```text
wliafe/star
├── README.md
├── images
│   ├── train
│   │   └── 0000.png
│   ├── validation
│   │   └── 0002.png
│   └── test
│       └── 0004.png
└── data
    ├── train-*.parquet
    ├── validation-*.parquet
    └── test-*.parquet
```

本地源数据中的 `val` 在 Hugging Face Dataset 中命名为 `validation`。

## 数据字段

每行表示一张图片及其场景图标注：

- `id`：图片文件名去除扩展名后的样本 ID。
- `image`：仓库相对路径，例如 `images/train/0000.png`。
- `width`、`height`：原图宽高。
- `polygons`：对象 polygon 列表；每个点为 `[x, y]`，保留原始坐标和顶点顺序。
- `labels`：与 `polygons` 一一对应的对象类别。
- `relations.subject_index`：关系主语在对象数组中的索引。
- `relations.object_index`：关系宾语在对象数组中的索引。
- `relations.predicate`：关系类别。

test split 只有图片，`polygons`、`labels` 和三个关系数组均为空。

## 下载并读取

`repo_type="dataset"` 是 `snapshot_download()` 的参数；`load_dataset()` 直接使用仓库 ID，不需要传入 `repo_type`。

```python
from pathlib import Path

from datasets import load_dataset
from huggingface_hub import snapshot_download


repo_dir = Path(
    snapshot_download(
        repo_id="wliafe/star",
        repo_type="dataset",
    )
)
dataset = load_dataset("wliafe/star")

sample = dataset["train"][0]
image_path = repo_dir / sample["image"]

print(sample["id"])
print(sample["image"])
print(image_path)
assert image_path.is_file()
```

`snapshot_download()` 返回仓库快照根目录，因此将它与 `sample["image"]` 拼接即可得到本地图片路径。不要直接把相对路径解释为当前工作目录下的文件。

### 使用 Pillow

```python
from PIL import Image


with Image.open(image_path) as image:
    image.load()
    print(image.size)
```

### 使用 OpenCV

```python
import cv2


image = cv2.imread(str(image_path), cv2.IMREAD_UNCHANGED)
if image is None:
    raise RuntimeError(f"无法读取图片：{image_path}")
print(image.shape)
```

## 固定数据版本

如果训练或评测需要可复现的数据版本，请为下载和 Dataset 加载指定同一个完整 commit revision：

```python
from pathlib import Path

from datasets import load_dataset
from huggingface_hub import snapshot_download


revision = "<full-commit-sha>"
repo_dir = Path(
    snapshot_download(
        repo_id="wliafe/star",
        repo_type="dataset",
        revision=revision,
    )
)
dataset = load_dataset("wliafe/star", revision=revision)

image_path = repo_dir / dataset["train"][0]["image"]
```

这样 Parquet 标注与原始图片始终来自同一个仓库版本。