Add dataset card

README.md

@@ -0,0 +1,104 @@
+---
+license: other
+license_name: mixed-upstream-licenses
+license_link: https://huggingface.co/datasets/royguw/mm-olmo-images#licensing-and-attribution
+pretty_name: mm-olmo images and video frames
+task_categories:
+- image-segmentation
+- object-detection
+- mask-generation
+tags:
+- grounding
+- referring-expression-segmentation
+- video-object-segmentation
+- image-mirror
+size_categories:
+- 1M<n<10M
+---
+
+# mm-olmo images and video frames
+
+A **pixel mirror**: the exact set of images and video frames that a grounding/segmentation
+training stack reads from outside its own annotation tree. It contains no annotations of its
+own — every mask, box, caption and vocabulary lives in the annotation parquets of the
+companion tree (`grounding_data`). This repo exists so that the annotations are actually
+usable by someone who does not have the original cluster filesystem.
+
+The file list is not a whole-dataset dump. It is exactly the files referenced by the
+training/eval parquets, which is 5–10× smaller than mirroring the upstream datasets whole
+(for example the BURST frames are 143,519 keyframes / 15.4 GiB, out of 1,574,936 files /
+171 GiB present in the source `TAO-Amodal` tree).
+
+## Contents
+
+| Group | Files | Size | Upstream source |
+|---|---|---|---|
+| `Objects365-2020` | 345,442 | 78.6 GiB | Objects365 v2 train images |
+| `SA-1B` | 71,077 | 56.8 GiB | SA-1B (high-instance-count subset) |
+| `BURST` | 143,519 | 15.4 GiB | TAO-Amodal frames (BURST annotates TAO videos) |
+| `LVOSv1` | 96,395 | 14.0 GiB | LVOS v1 |
+| `OVIS` | 53,857 | 12.6 GiB | OVIS |
+| `pixmo_images` | 56,831 | 10.0 GiB | PixMo pointing images |
+| `Ref-YT-VOS` | 100,074 | 8.5 GiB | Refer-YouTube-VOS |
+| `LV-VIS` | 88,577 | 8.4 GiB | LV-VIS |
+| `YT-VIS` | 30,452 | 7.4 GiB | YouTube-VIS (JPEG dirs + train `.mp4`s) |
+| `ADE20k` | 55,151 | 6.0 GiB | ADE20K, Supervisely format (`img/` + `ann/` + parquet caches) |
+| `ReVOS` | 35,161 | 5.8 GiB | ReVOS |
+| `Ref-DAVIS17` | 6,208 | 895 MiB | Refer-DAVIS17 |
+| **Total** | **1,082,744** | **224.5 GiB** | 52 tar shards |
+
+## Layout
+
+```
+shards/<group>/<group>-00000.tar   # ~5 GiB each; tar members are paths relative to the mm-olmo root
+index/<group>.parquet              # path -> shard, size  (for fetching selectively)
+```
+
+Packed as tar shards rather than loose files because the Hub asks for fewer than 100k files
+per repo and at most 10k entries per folder; 1.08M loose images would violate both. Note
+this means the dataset viewer does not apply — the payload is a file tree, not tabular rows.
+
+## Reconstructing the tree
+
+Tar members are relative paths, so extracting every shard into one directory rebuilds the
+original layout:
+
+```bash
+hf download royguw/mm-olmo-images --repo-type dataset --local-dir ./dl
+MM_ROOT=/your/path/to/mm-olmo
+for t in ./dl/shards/*/*.tar; do tar -xf "$t" -C "$MM_ROOT"; done
+```
+
+You then get `$MM_ROOT/img_datasets/…`, `$MM_ROOT/images/ADE20k/…`,
+`$MM_ROOT/torch_datasets/pixmo_images/…` and `$MM_ROOT/video_datasets/…`.
+
+To grab one subset only, read its `index/<group>.parquet`, take the distinct `shard`
+values, and download just those:
+
+```python
+import pyarrow.parquet as pq
+from huggingface_hub import hf_hub_download
+
+idx = pq.read_table(hf_hub_download("royguw/mm-olmo-images", "index/ADE20k.parquet",
+                                    repo_type="dataset")).to_pydict()
+for shard in dict.fromkeys(idx["shard"]):
+    hf_hub_download("royguw/mm-olmo-images", shard, repo_type="dataset")
+```
+
+The annotation parquets that reference these files store absolute paths, so after extracting
+you will need to rewrite their stored prefix to `$MM_ROOT`.
+
+## Licensing and attribution
+
+This is a **redistribution mirror of third-party datasets**, assembled only to make the
+companion annotations reproducible. It is not a new dataset and no new license is claimed
+over the pixels.
+
+Each group above remains under the license of its upstream source, and those licenses
+differ — several are research/non-commercial only, and some require agreeing to upstream
+terms before use. **Before using any group, check that group's upstream license and comply
+with it.** If you are an author or rights-holder of any upstream dataset here and want a
+subset removed, open a discussion on this repo and it will be taken down.
+
+No attempt has been made to relicense, and the per-group provenance table above is provided
+precisely so that each source stays traceable.