wliafe commited on
Commit
9158e93
·
verified ·
1 Parent(s): 007bc81

Add STAR dataset card

Browse files
Files changed (1) hide show
  1. README.md +131 -0
README.md ADDED
@@ -0,0 +1,131 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ pretty_name: STAR Relationship
3
+ tags:
4
+ - scene-graph-generation
5
+ - remote-sensing
6
+ - polygon-annotation
7
+ ---
8
+
9
+ # STAR Relationship
10
+
11
+ 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:
12
+
13
+ - 图片以普通文件形式位于 `images/`。
14
+ - train、validation 和 test 的结构化标注以 Parquet 保存。
15
+ - 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。
16
+
17
+ 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。
18
+
19
+ ## 仓库结构
20
+
21
+ ```text
22
+ wliafe/star
23
+ ├── README.md
24
+ ├── images
25
+ │ ├── train
26
+ │ │ └── 0000.png
27
+ │ ├── validation
28
+ │ │ └── 0002.png
29
+ │ └── test
30
+ │ └── 0004.png
31
+ └── data
32
+ ├── train-*.parquet
33
+ ├── validation-*.parquet
34
+ └── test-*.parquet
35
+ ```
36
+
37
+ 本地源数据中的 `val` 在 Hugging Face Dataset 中命名为 `validation`。
38
+
39
+ ## 数据字段
40
+
41
+ 每行表示一张图片及其场景图标注:
42
+
43
+ - `id`:图片文件名去除扩展名后的样本 ID。
44
+ - `image`:仓库相对路径,例如 `images/train/0000.png`。
45
+ - `width`、`height`:原图宽高。
46
+ - `polygons`:对象 polygon 列表;每个点为 `[x, y]`,保留原始坐标和顶点顺序。
47
+ - `labels`:与 `polygons` 一一对应的对象类别。
48
+ - `relations.subject_index`:关系主语在对象数组中的索引。
49
+ - `relations.object_index`:关系宾语在对象数组中的索引。
50
+ - `relations.predicate`:关系类别。
51
+
52
+ test split 只有图片,`polygons`、`labels` 和三个关系数组均为空。
53
+
54
+ ## 下载并读取
55
+
56
+ `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`。
57
+
58
+ ```python
59
+ from pathlib import Path
60
+
61
+ from datasets import load_dataset
62
+ from huggingface_hub import snapshot_download
63
+
64
+
65
+ repo_dir = Path(
66
+ snapshot_download(
67
+ repo_id="wliafe/star",
68
+ repo_type="dataset",
69
+ )
70
+ )
71
+ dataset = load_dataset("wliafe/star")
72
+
73
+ sample = dataset["train"][0]
74
+ image_path = repo_dir / sample["image"]
75
+
76
+ print(sample["id"])
77
+ print(sample["image"])
78
+ print(image_path)
79
+ assert image_path.is_file()
80
+ ```
81
+
82
+ `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。
83
+
84
+ ### 使用 Pillow
85
+
86
+ ```python
87
+ from PIL import Image
88
+
89
+
90
+ with Image.open(image_path) as image:
91
+ image.load()
92
+ print(image.size)
93
+ ```
94
+
95
+ ### 使用 OpenCV
96
+
97
+ ```python
98
+ import cv2
99
+
100
+
101
+ image = cv2.imread(str(image_path), cv2.IMREAD_UNCHANGED)
102
+ if image is None:
103
+ raise RuntimeError(f"无法读取图片:{image_path}")
104
+ print(image.shape)
105
+ ```
106
+
107
+ ## 固定数据版本
108
+
109
+ 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:
110
+
111
+ ```python
112
+ from pathlib import Path
113
+
114
+ from datasets import load_dataset
115
+ from huggingface_hub import snapshot_download
116
+
117
+
118
+ revision = "<full-commit-sha>"
119
+ repo_dir = Path(
120
+ snapshot_download(
121
+ repo_id="wliafe/star",
122
+ repo_type="dataset",
123
+ revision=revision,
124
+ )
125
+ )
126
+ dataset = load_dataset("wliafe/star", revision=revision)
127
+
128
+ image_path = repo_dir / dataset["train"][0]["image"]
129
+ ```
130
+
131
+ 这样 Parquet 标注与原始图片始终来自同一个仓库版本。