Add OIV6 dataset description

#1
by wliafe - opened
Files changed (1) hide show
  1. README.md +106 -0
README.md CHANGED
@@ -682,3 +682,109 @@ dataset_info:
682
  download_size: 40106623995
683
  dataset_size: 40101849835
684
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
682
  download_size: 40106623995
683
  dataset_size: 40101849835
684
  ---
685
+
686
+ # Open Images V6 Relationships
687
+
688
+ OIV6 是基于 Open Images V6 的视觉关系检测数据集,包含 133,503 张图片、
689
+ 601 个对象前景类别和 30 个关系谓词前景类别。图片字节直接嵌入 Parquet,
690
+ 可通过 Hugging Face `Image` feature 解码。
691
+
692
+ ## 数据集规模
693
+
694
+ | Split | 图片 | 对象 | 关系 |
695
+ | --- | ---: | ---: | ---: |
696
+ | `train` | 126,368 | 512,259 | 348,560 |
697
+ | `validation` | 1,813 | 6,386 | 4,951 |
698
+ | `test` | 5,322 | 19,284 | 14,403 |
699
+ | **总计** | **133,503** | **537,929** | **367,914** |
700
+
701
+ 三个 split 互不重叠。每个样本都包含至少一个对象和一条关系。
702
+
703
+ ## 加载
704
+
705
+ ```python
706
+ from datasets import load_dataset
707
+
708
+ dataset = load_dataset("wliafe/OIV6")
709
+ sample = dataset["train"][0]
710
+
711
+ image = sample["image"] # PIL.Image.Image
712
+ print(sample["id"], image.size)
713
+ ```
714
+
715
+ 图片已嵌入 Parquet,不需要额外下载或拼接图片目录。
716
+
717
+ ## 数据字段
718
+
719
+ | 字段 | 类型 | 说明 |
720
+ | --- | --- | --- |
721
+ | `id` | `string` | Open Images 图片 ID |
722
+ | `image` | `Image` | 可直接解码的嵌入式 JPEG |
723
+ | `width` | `int32` | JPEG 实际宽度,单位为像素 |
724
+ | `height` | `int32` | JPEG 实际高度,单位为像素 |
725
+ | `boxes` | `List[[float32; 4]]` | 与对象平行的 `[x1, y1, x2, y2]` 边界框 |
726
+ | `labels` | `List[ClassLabel]` | 与 `boxes` 平行的对象类别 |
727
+ | `relations.subject_index` | `List[int64]` | 关系主语在当前对象数组中的索引 |
728
+ | `relations.object_index` | `List[int64]` | 关系宾语在当前对象数组中的索引 |
729
+ | `relations.predicate` | `List[ClassLabel]` | 关系谓词类别 |
730
+
731
+ `boxes` 和 `labels` 长度相同。三个关系数组也具有相同长度;相同位置的主语索引、
732
+ 宾语索引和谓词共同表示一条有向关系。
733
+
734
+ ## 类别与关系名称
735
+
736
+ 对象和谓词 taxonomy 均在索引 `0` 保留 `__background__`:
737
+
738
+ - 对象前景类别编号为 `1`–`601`。
739
+ - 谓词前景类别编号为 `1`–`30`。
740
+ - `subject_index` 和 `object_index` 是当前样本对象数组的零基位置,不是类别 ID。
741
+
742
+ ```python
743
+ features = dataset["train"].features
744
+ object_names = features["labels"].feature.names
745
+ predicate_names = features["relations"]["predicate"].feature.names
746
+
747
+ sample = dataset["train"][0]
748
+ for subject, object_, predicate in zip(
749
+ sample["relations"]["subject_index"],
750
+ sample["relations"]["object_index"],
751
+ sample["relations"]["predicate"],
752
+ ):
753
+ print(
754
+ object_names[sample["labels"][subject]],
755
+ predicate_names[predicate],
756
+ object_names[sample["labels"][object_]],
757
+ )
758
+ ```
759
+
760
+ ## 坐标约定
761
+
762
+ `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
763
+ 边界内。`width` 和 `height` 与解码后 `image` 的尺寸一致。边界框表示对象检测
764
+ 区域,不是实例分割轮廓。
765
+
766
+ ## 使用限制
767
+
768
+ - 对象和关系类别呈长尾分布,模型结果可能被高频类别主导。
769
+ - 标注可能包含遗漏、歧义或类别噪声。
770
+ - 本仓库不重新授予原始图片版权;使用者应遵守 Open Images 的许可与使用要求。
771
+ - 比较模型结果时,应确认使用相同的 taxonomy、background 编号和 split。
772
+
773
+ ## 引用
774
+
775
+ 使用本数据集时,请引用 Open Images:
776
+
777
+ ```bibtex
778
+ @article{kuznetsova2020open,
779
+ title={The Open Images Dataset V4: Unified Image Classification,
780
+ Object Detection, and Visual Relationship Detection at Scale},
781
+ author={Kuznetsova, Alina and Rom, Hassan and Alldrin, Neil and
782
+ Uijlings, Jasper and Krasin, Ivan and Pont-Tuset, Jordi and
783
+ Kamali, Shahab and Popov, Stefan and Malloci, Matteo and
784
+ Kolesnikov, Alexander and Duerig, Tom and Ferrari, Vittorio},
785
+ journal={International Journal of Computer Vision},
786
+ volume={128},
787
+ pages={1956--1981},
788
+ year={2020}
789
+ }
790
+ ```