Upload README.md with huggingface_hub

README.md

@@ -1,28 +1,70 @@
 ---
-dataset_info:
-  features:
-  - name: image
-    dtype: image
-  - name: mask
-    dtype: image
-  - name: instance_id
-    dtype: int32
-  - name: subset
-    dtype: string
-  splits:
-  - name: comp
-    num_bytes: 5644136440
-    num_examples: 5883
-  - name: natural
-    num_bytes: 778747388
-    num_examples: 830
-  download_size: 6423302154
-  dataset_size: 6422883828
-configs:
-- config_name: default
-  data_files:
-  - split: comp
-    path: data/comp-*
-  - split: natural
-    path: data/natural-*
+license: cc-by-nc-4.0
+pretty_name: HIM-2K
+task_categories:
+- image-segmentation
+tags:
+- human-matting
+- instance-matting
+- alpha-matting
+- image-matting
+size_categories:
+- 1K<n<10K
 ---
+
+# HIM-2K (Human Instance Matting)
+
+HIM-2K is a **human instance matting** benchmark introduced with **InstMatt**
+(["Human Instance Matting via Mutual Guidance and Multi-Instance Refinement", Sun et al., CVPR 2022](https://openaccess.thecvf.com/content/CVPR2022/html/Sun_Human_Instance_Matting_via_Mutual_Guidance_and_Multi-Instance_Refinement_CVPR_2022_paper.html)).
+Unlike standard human matting, it provides a **separate alpha matte for every human instance** in an image,
+so a single photo containing multiple people yields multiple ground-truth alpha mattes.
+
+Source: [nowsyn/InstMatt](https://github.com/nowsyn/InstMatt).
+
+## Dataset structure
+
+Because instance matting has a variable number of alpha mattes per image, this dataset uses
+**one row per (image, instance-alpha) pair**. All rows in every split share the same schema:
+
+| column        | type            | description                                                        |
+|---------------|-----------------|--------------------------------------------------------------------|
+| `image`       | `Image`         | The shared RGB photo. Repeated across the instances it contains.   |
+| `mask`        | `Image`         | The alpha matte for **one** human instance in that image.          |
+| `instance_id` | `int32`         | Index of the instance within its image (from the alpha filename).  |
+| `subset`      | `string`        | `"comp"` or `"natural"` (matches the split name).                  |
+
+An image with N humans appears in N rows, all sharing the same `image` but each carrying a
+different `mask` / `instance_id`.
+
+## Splits
+
+The dataset's own subset structure is used as the splits (there is no separate train/test split):
+
+| split     | images | instance rows |
+|-----------|--------|---------------|
+| `comp`    | 1680   | 5883          |
+| `natural` | 320    | 830           |
+
+- **`comp`** — composited images.
+- **`natural`** — real natural images with ground-truth per-instance alphas.
+
+Note: the original archive also ships an `images/natural_wo_gt` folder (100 natural images
+*without* ground-truth alphas). Those images have no masks and therefore cannot fit the
+`image` + `mask` schema, so they are **excluded** from this dataset.
+
+## License
+
+Released under **CC BY-NC 4.0** (academic / non-commercial). The upstream InstMatt repository
+does not state an explicit license; this is a conservative choice for an academic matting benchmark.
+Please cite the InstMatt paper if you use this data.
+
+## Citation
+
+```bibtex
+@inproceedings{sun2022instmatt,
+  title     = {Human Instance Matting via Mutual Guidance and Multi-Instance Refinement},
+  author    = {Sun, Yanan and Tang, Chi-Keung and Tai, Yu-Wing},
+  booktitle = {CVPR},
+  year      = {2022}
+}
+```