wliafe commited on
Commit
0b58b8d
·
verified ·
1 Parent(s): f82c671

Add dataset description

Browse files
Files changed (1) hide show
  1. README.md +147 -0
README.md CHANGED
@@ -457,3 +457,150 @@ configs:
457
  - split: test
458
  path: data/test-*
459
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
457
  - split: test
458
  path: data/test-*
459
  ---
460
+
461
+ # VG150
462
+
463
+ VG150 是从 [Visual Genome](https://visualgenome.org/) 派生的场景图生成
464
+ (Scene Graph Generation, SGG)基准数据集,包含 150 个对象类别、50 个
465
+ 关系谓词类别和 200 个属性类别。图片直接嵌入 Parquet。
466
+
467
+ 该数据集适用于场景图生成、场景图检测、视觉关系检测和相关多模态研究。
468
+
469
+ ## 数据集规模
470
+
471
+ | Split | 图片数量 | 用途 |
472
+ | --- | ---: | --- |
473
+ | `train` | 57,723 | 模型训练 |
474
+ | `validation` | 5,000 | 验证与模型选择 |
475
+ | `test` | 26,446 | 最终评估 |
476
+ | **总计** | **89,169** | 仅包含同时具有对象和关系标注的图片 |
477
+
478
+ 所有 split 均包含完整的对象、属性和关系标注。
479
+
480
+ ## 加载
481
+
482
+ ```python
483
+ from datasets import load_dataset
484
+
485
+ dataset = load_dataset("wliafe/VG150")
486
+ print(dataset)
487
+
488
+ sample = dataset["train"][0]
489
+ image = sample["image"] # PIL.Image.Image
490
+ image.show()
491
+ ```
492
+
493
+ 图片字节已经嵌入 Parquet,因此不需要额外下载或拼接 Visual Genome 图片目录。
494
+ `image` 使用 Hugging Face `Image` feature,默认读取时会解码为 Pillow 图像。
495
+
496
+ 如需取得类别名称,可以读取 `ClassLabel` feature:
497
+
498
+ ```python
499
+ features = dataset["train"].features
500
+
501
+ object_names = features["labels"].feature.names
502
+ predicate_names = features["relations"]["predicate"].feature.names
503
+ attribute_names = features["attributes"].feature.feature.names
504
+
505
+ sample = dataset["train"][0]
506
+ object_labels = [object_names[index] for index in sample["labels"]]
507
+ predicate_labels = [
508
+ predicate_names[index]
509
+ for index in sample["relations"]["predicate"]
510
+ ]
511
+ attribute_labels = [
512
+ [attribute_names[index] for index in object_attributes]
513
+ for object_attributes in sample["attributes"]
514
+ ]
515
+ ```
516
+
517
+ ## 数据字段
518
+
519
+ | 字段 | 类型 | 说明 |
520
+ | --- | --- | --- |
521
+ | `id` | `string` | Visual Genome 图片 ID |
522
+ | `image` | `Image` | 嵌入 Parquet 并可直接解码的图片 |
523
+ | `width` | `int32` | 实际 JPEG 宽度,单位为像素 |
524
+ | `height` | `int32` | 实际 JPEG 高度,单位为像素 |
525
+ | `boxes` | `List[[float32; 4]]` | 与对象平行的 `[x1, y1, x2, y2]` 边界框 |
526
+ | `labels` | `List[ClassLabel]` | 与 `boxes` 平行的对象类别 |
527
+ | `attributes` | `List[List[ClassLabel]]` | 每个对象对应的零个或多个属性 |
528
+ | `relations.subject_index` | `List[int64]` | 每条关系的主语对象索引 |
529
+ | `relations.object_index` | `List[int64]` | 每条关系的宾语对象索引 |
530
+ | `relations.predicate` | `List[ClassLabel]` | 每条关系的谓词类别 |
531
+
532
+ `boxes`、`labels` 和 `attributes` 的外层长度相同。三个关系数组也具有相同
533
+ 长度;相同位置的主语索引、宾语索引和谓词共同构成一条有向关系:
534
+
535
+ ```python
536
+ sample = dataset["train"][0]
537
+
538
+ for subject, object_, predicate in zip(
539
+ sample["relations"]["subject_index"],
540
+ sample["relations"]["object_index"],
541
+ sample["relations"]["predicate"],
542
+ ):
543
+ print(
544
+ object_names[sample["labels"][subject]],
545
+ predicate_names[predicate],
546
+ object_names[sample["labels"][object_]],
547
+ )
548
+ ```
549
+
550
+ ## 类别编号
551
+
552
+ 对象、谓词和属性 taxonomy 均在索引 `0` 保留 `__background__`:
553
+
554
+ - 对象前景类别编号为 `1`–`150`。
555
+ - 谓词前景类别编号为 `1`–`50`。
556
+ - 属性前景类别编号为 `1`–`200`。
557
+
558
+ 没有属性的对象对应空列表,而不是包含 `__background__` 的列表。
559
+
560
+ ## 边界框与图片尺寸
561
+
562
+ `boxes` 使用实际图片像素坐标下的 `[x1, y1, x2, y2]` 格式,坐标位于图片
563
+ 边界内。`width` 和 `height` 与 `image` 的实际尺寸一致。关系端点使用当前样本
564
+ 对象数组中的零基索引。
565
+
566
+ 本仓库提供可直接加载的 Parquet 数据,不包含原始 H5 或 JSON 文件。对象、关系和
567
+ 属性标注可能继承 Visual Genome 中的类别偏差、长尾分布、歧义和错误。
568
+
569
+ ## 使用限制
570
+
571
+ - VG150 只保留高频类别,不能代表 Visual Genome 的完整开放词汇分布。
572
+ - 关系类别分布高度不均衡,模型结果可能被少数高频谓词主导。
573
+ - `boxes` 是对象边界框,不是实例分割轮廓。
574
+ - 本仓库不重新授予原始图片版权。使用者应同时遵守 Visual Genome 的使用条款,
575
+ 并确认原始图片适用于自己的研究或发布场景。
576
+ - 比较论文结果时,应确认对方使用相同的 VG150 taxonomy、过滤规则和 split。
577
+
578
+ ## 引用
579
+
580
+ 使用本数据集时,请引用 Visual Genome:
581
+
582
+ ```bibtex
583
+ @article{krishna2017visual,
584
+ title={Visual Genome: Connecting Language and Vision Using Crowdsourced Dense Image Annotations},
585
+ author={Krishna, Ranjay and Zhu, Yuke and Groth, Oliver and Johnson, Justin and
586
+ Hata, Kenji and Kravitz, Joshua and Chen, Stephanie and
587
+ Kalantidis, Yannis and Li, Li-Jia and Shamma, David A. and
588
+ Bernstein, Michael S. and Fei-Fei, Li},
589
+ journal={International Journal of Computer Vision},
590
+ volume={123},
591
+ number={1},
592
+ pages={32--73},
593
+ year={2017}
594
+ }
595
+ ```
596
+
597
+ 标准 VG150 场景图划分也常用于 Neural Motifs:
598
+
599
+ ```bibtex
600
+ @inproceedings{zellers2018neural,
601
+ title={Neural Motifs: Scene Graph Parsing with Global Context},
602
+ author={Zellers, Rowan and Yatskar, Mark and Thomson, Sam and Choi, Yejin},
603
+ booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
604
+ year={2018}
605
+ }
606
+ ```