Document the ReCon1M dataset

Document schema, image path resolution, splits, source ID remapping, and citation while preserving the generated YAML front matter (SHA-256: 63ad10bc99e3a7c4fe247126f74de7f0823b431f397587989c1fde8fd092bc89).

README.md

@@ -176,3 +176,124 @@ configs:
   - split: test
     path: data/test-*
 ---
+
+
+# ReCon1M
+
+ReCon1M is a remote-sensing scene graph generation dataset derived from
+FAIR1M. This repository packages the local release as structured Hugging Face
+Parquet annotations while keeping PNG images as ordinary repository files.
+
+The packaged release contains **22,262 images**,
+**60 object classes**, and
+**59 predicate classes**:
+
+| Split | Images |
+| --- | ---: |
+| `train` | 11,131 |
+| `validation` | 3,710 |
+| `test` | 7,421 |
+
+These counts describe the files in this packaged release. They differ from the
+counts in the paper abstract, which may describe another ReCon1M release.
+
+## Repository layout
+
+```text
+images/<split>/<first-two-ID-characters>/<id>.png
+data/...
+README.md
+```
+
+The `image` column is a POSIX repository-relative path such as
+`images/train/09/09868.png`. It contains neither image bytes nor an automatic
+Pillow object. Most images use `.png`; the three JPEG-encoded source images use
+their corrected `.jpeg` filenames: `00024.jpeg`, `00033.jpeg`, and
+`00037.jpeg`.
+
+## Schema
+
+- `id`: original zero-padded image identifier.
+- `image`: repository-relative PNG path.
+- `width`, `height`: PNG dimensions.
+- `image_source`, `gsd`: metadata from the object annotation header.
+- `polygons`: one four-point polygon per object. Coordinates and vertex order
+  are preserved; polygons are not converted to bounding boxes.
+- `labels`: object `ClassLabel` values.
+- `difficult`: original per-object difficulty integers.
+- `relations.subject_index`, `relations.object_index`: zero-based indices into
+  `labels` and `polygons`.
+- `relations.predicate`: predicate `ClassLabel` values.
+
+The source object and predicate vocabularies use one-based IDs. This package
+maps them to zero-based Hugging Face `ClassLabel` indices without changing
+their vocabulary order. No `attributes` field is added.
+
+## Download and open images
+
+Install `datasets`, `huggingface_hub`, and Pillow, then download the repository
+snapshot and load its Parquet annotations:
+
+```python
+from pathlib import Path
+
+from datasets import load_dataset
+from huggingface_hub import snapshot_download
+from PIL import Image
+
+
+repo_dir = Path(snapshot_download("wliafe/recon1m", repo_type="dataset"))
+dataset = load_dataset("wliafe/recon1m")
+
+sample = dataset["train"][0]
+image_path = repo_dir / sample["image"]
+
+with Image.open(image_path) as image:
+    image.load()
+    print(image.size)
+```
+
+OpenCV can read the same resolved path:
+
+```python
+import cv2
+
+
+image_bgr = cv2.imread(str(image_path), cv2.IMREAD_COLOR)
+if image_bgr is None:
+    raise RuntimeError(f"Failed to read {image_path}")
+```
+
+The complete snapshot includes roughly 32.6 GB of PNG files. Calling only
+`load_dataset()` downloads the Parquet annotations, not all image files.
+
+For reproducible loading, pass the same revision to both calls:
+
+```python
+revision = "<commit-sha>"
+repo_dir = Path(
+    snapshot_download(
+        "wliafe/recon1m",
+        repo_type="dataset",
+        revision=revision,
+    )
+)
+dataset = load_dataset("wliafe/recon1m", revision=revision)
+```
+
+## Citation
+
+If you use ReCon1M, cite the original paper:
+
+```bibtex
+@article{sun2024recon1m,
+  title={ReCon1M: A Large-scale Benchmark Dataset for Relation Comprehension in Remote Sensing Imagery},
+  author={Sun, Xian and Yan, Qiwei and Deng, Chubo and Liu, Chenglong and Jiang, Yi and Hou, Zhongyan and Lu, Wanxuan and Yao, Fanglong and Liu, Xiaoyu and Hao, Lingxiang and Yu, Hongfeng},
+  journal={arXiv preprint arXiv:2406.06028},
+  year={2024}
+}
+```
+
+See the [ReCon1M paper](https://arxiv.org/abs/2406.06028) for the dataset
+methodology. Use of the images and annotations remains subject to the terms of
+the original ReCon1M and FAIR1M releases.