Document the GQA200 dataset

README.md

@@ -350,3 +350,82 @@ configs:
   - split: validation
     path: data/validation-*
 ---
+# GQA200
+
+GQA200 is the scene-graph-generation benchmark subset of
+[GQA](https://cs.stanford.edu/people/dorarad/gqa/index.html). This repository
+uses the standard GQA200 taxonomy with 200 foreground object classes
+and 100 foreground predicate classes.
+
+## Splits
+
+| Split | Images | Source |
+| --- | ---: | --- |
+| `train` | 57,623 | Standard GQA200 Train |
+| `validation` | 8,209 | Standard GQA200 Test |
+
+The standard GQA200 Test annotations are intentionally exposed as
+`validation`. This repository does not define a separate `test` split.
+
+## Fields
+
+- `id`: GQA image identifier.
+- `image`: image embedded in Parquet and decoded by `datasets` as a PIL image.
+- `width`, `height`: image dimensions.
+- `location`, `weather`: nullable strings copied from the original GQA scene
+  graph metadata.
+- `boxes`: object boxes in `[x1, y1, x2, y2]` order.
+- `labels`: GQA200 object `ClassLabel` IDs.
+- `relations`: parallel `subject_index`, `object_index`, and predicate arrays.
+
+Index 0 is `__background__` in both ClassLabel vocabularies. Foreground object
+IDs are 1–200 and foreground predicate IDs are 1–100. Attributes and the full
+raw GQA taxonomy are not included.
+
+Boxes are clipped to image bounds. Objects with zero area after clipping are
+removed, remaining object indices are compacted, and relations touching removed
+objects are discarded.
+
+## Loading
+
+```python
+from datasets import load_dataset
+
+dataset = load_dataset("wliafe/GQA200")
+sample = dataset["train"][0]
+
+image = sample["image"]  # PIL.Image.Image
+object_name = dataset["train"].features["labels"].feature.int2str(
+    sample["labels"][0]
+)
+predicate_name = (
+    dataset["train"]
+    .features["relations"]["predicate"]
+    .feature.int2str(sample["relations"]["predicate"][0])
+)
+```
+
+`location` and `weather` may be `None` when the original scene graph omits the
+field.
+
+## Sources and citation
+
+The original format is documented on the
+[GQA download page](https://cs.stanford.edu/people/dorarad/gqa/download.html).
+The GQA200 benchmark split and taxonomy follow the SHA-GCL evaluation setup.
+
+```bibtex
+@inproceedings{hudson2019gqa,
+  title={GQA: A New Dataset for Real-World Visual Reasoning and Compositional Question Answering},
+  author={Hudson, Drew A. and Manning, Christopher D.},
+  booktitle={CVPR},
+  year={2019}
+}
+
+@inproceedings{dong2022stacked,
+  title={Stacked Hybrid-Attention and Group Collaborative Learning for Unbiased Scene Graph Generation},
+  author={Dong, Xingning and Gan, Tian and Song, Xianjing and Wu, Jinhui and Cheng, Yuan and Nie, Liqiang},
+  booktitle={CVPR},
+  year={2022}
+}
+```