Add PSG dataset description

#1
by wliafe - opened
Files changed (1) hide show
  1. README.md +120 -0
README.md CHANGED
@@ -247,3 +247,123 @@ configs:
247
  - split: validation
248
  path: data/validation-*
249
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
247
  - split: validation
248
  path: data/validation-*
249
  ---
250
+
251
+ # Panoptic Scene Graph
252
+
253
+ PSG(Panoptic Scene Graph)是面向全景场景图生成的数据集。每个场景图节点不仅包含对象类别和边界框,还通过 panoptic segmentation mask 与像素级区域对应。数据集同时覆盖 thing 和 stuff 类别。
254
+
255
+ 本仓库将 RGB 图片和 panoptic PNG 直接嵌入 Parquet。通过 Hugging Face `Image` feature 加载后,两者都会解码为 Pillow 图像,不需要额外下载图片目录。
256
+
257
+ ## 数据集规模
258
+
259
+ | Split | 样本 | 对象 / segments | 关系 |
260
+ | --- | ---: | ---: | ---: |
261
+ | `train` | 46,563 | 513,243 | 261,666 |
262
+ | `validation` | 2,186 | 24,910 | 13,705 |
263
+ | **总计** | **48,749** | **538,153** | **275,371** |
264
+
265
+ `validation` 对应完整 PSG 发布版中的官方 test split。完整发布版带有 test ground truth,因此这里保留其对象、mask 和关系标注,并改名为 Hugging Face 常用的 `validation`。部分样本没有关系,其关系数组为空。
266
+
267
+ PSG 中有少量不同样本引用同一张 COCO 图片;`id` 是 PSG 样本 ID,不应使用图片文件名替代。
268
+
269
+ ## 加载
270
+
271
+ ```python
272
+ from datasets import load_dataset
273
+
274
+ dataset = load_dataset("wliafe/PSG")
275
+ sample = dataset["train"][0]
276
+
277
+ image = sample["image"] # PIL.Image.Image
278
+ panoptic_mask = sample["panoptic_mask"] # PIL.Image.Image
279
+ print(sample["id"], image.size, panoptic_mask.size)
280
+ ```
281
+
282
+ ## 字段
283
+
284
+ | 字段 | 类型 | 说明 |
285
+ | --- | --- | --- |
286
+ | `id` | `string` | PSG 样本 ID |
287
+ | `image` | `Image` | 嵌入式 RGB 图片;源文件主要是 JPEG,少量 `.jpg` 实际为 PNG |
288
+ | `panoptic_mask` | `Image` | 嵌入式 RGB 编码 panoptic PNG |
289
+ | `width` / `height` | `int32` | 解码后 RGB 图片的实际像素尺寸 |
290
+ | `boxes` | `List[[float32; 4]]` | 与 segments 平行的 `[x1, y1, x2, y2]` 边界框 |
291
+ | `labels` | `List[ClassLabel]` | 与 segments 平行的对象类别 |
292
+ | `segments.id` | `List[int64]` | panoptic mask 中的 segment ID |
293
+ | `segments.area` | `List[int64]` | segment 的标注面积 |
294
+ | `segments.iscrowd` | `List[bool]` | COCO crowd 标记 |
295
+ | `segments.isthing` | `List[bool]` | thing/stuff 标记 |
296
+ | `relations.subject_index` | `List[int64]` | 主语在当前 segment 数组中的零基索引 |
297
+ | `relations.object_index` | `List[int64]` | 宾语在当前 segment 数组中的零基索引 |
298
+ | `relations.predicate` | `List[ClassLabel]` | 有向关系的谓词类别 |
299
+
300
+ `boxes`、`labels` 和四个 `segments` 数组长度相同。同一位置表示同一个场景图节点。三个 `relations` 数组长度也相同,同一位置共同表示一条有向关系。
301
+
302
+ ## Panoptic mask
303
+
304
+ panoptic PNG 使用 RGB 三通道编码整数 segment ID:
305
+
306
+ ```python
307
+ import numpy as np
308
+
309
+ rgb = np.asarray(sample["panoptic_mask"].convert("RGB"), dtype=np.int64)
310
+ segment_id_map = (
311
+ rgb[..., 0]
312
+ + 256 * rgb[..., 1]
313
+ + 256 * 256 * rgb[..., 2]
314
+ )
315
+
316
+ first_segment_id = sample["segments"]["id"][0]
317
+ first_segment_mask = segment_id_map == first_segment_id
318
+ print(first_segment_mask.sum())
319
+ ```
320
+
321
+ mask 中的非零 segment ID 与 `segments.id` 对应。背景像素的 ID 为 0,这与 taxonomy 中保留的 `__background__` 类别编号是两个不同概念。
322
+
323
+ ## 类别与关系
324
+
325
+ 对象和谓词 taxonomy 都在索引 `0` 保留 `__background__`:
326
+
327
+ - 对象前景类别编号为 `1`–`133`,依次对应 80 个 thing 类和 53 个 stuff 类。
328
+ - 谓词前景类别编号为 `1`–`56`。
329
+ - `subject_index` 和 `object_index` 是样本内部 segment 数组的位置,不是类别 ID。
330
+
331
+ ```python
332
+ features = dataset["train"].features
333
+ object_names = features["labels"].feature.names
334
+ predicate_names = features["relations"]["predicate"].feature.names
335
+
336
+ for subject, object_, predicate in zip(
337
+ sample["relations"]["subject_index"],
338
+ sample["relations"]["object_index"],
339
+ sample["relations"]["predicate"],
340
+ ):
341
+ print(
342
+ object_names[sample["labels"][subject]],
343
+ predicate_names[predicate],
344
+ object_names[sample["labels"][object_]],
345
+ )
346
+ ```
347
+
348
+ ## 坐标约定
349
+
350
+ `boxes` 使用实际图片像素坐标下的 `[x1, y1, x2, y2]` 格式,并位于图片边界内。`width` 和 `height` 与解码后的 RGB 图片及 panoptic mask 尺寸一致。
351
+
352
+ ## 使用限制
353
+
354
+ - 类别和谓词呈长尾分布,评估结果可能受高频类别主导。
355
+ - 场景图和分割标注可能包含遗漏、歧义或类别噪声。
356
+ - 本仓库不重新授予原始 COCO 图片版权;使用者应遵守 COCO 与 PSG 的许可和使用要求。
357
+ - 比较模型结果时,应确认采用相同 split、taxonomy 和 background 编号。
358
+
359
+ ## 引用
360
+
361
+ ```bibtex
362
+ @inproceedings{yang2022panoptic,
363
+ title={Panoptic Scene Graph Generation},
364
+ author={Yang, Jingkang and Ang, Yi Zhe and Guo, Zujin and Zhou, Kaiyang
365
+ and Zhang, Wayne and Liu, Ziwei},
366
+ booktitle={European Conference on Computer Vision},
367
+ year={2022}
368
+ }
369
+ ```