README.md

---
license: other
license_name: mixed-upstream-licenses
license_link: https://huggingface.co/datasets/royguw/grounding-data#licensing-and-attribution
pretty_name: grounding_data - annotations and images
task_categories:
- image-segmentation
- object-detection
- mask-generation
- image-to-text
tags:
- grounding
- referring-expression-segmentation
- video-object-segmentation
- panoptic-segmentation
- grounded-captioning
size_categories:
- 1M<n<10M
---

# grounding_data

The annotation tree behind a grounding/segmentation training stack, plus the images and
video frames that live alongside it. Unlike the companion
[`royguw/mm-olmo-images`](https://huggingface.co/datasets/royguw/mm-olmo-images) — which is
pixels only — **this repo carries the annotations**: the parquet caches, JSON label files
and vocabularies under each dataset's `cache/` directory, which is where masks, boxes,
referring expressions, captions and category vocabularies actually live.

9,331,994 files / 1.20 TiB, packed as 255 tar shards across 34 groups.

## Layout

```
shards/<group>/<group>-00000.tar   # ~5 GiB each; members are paths relative to the grounding_data root
index/<group>.parquet              # path -> shard, size  (for fetching selectively)
```

Tar shards rather than loose files because the Hub caps entries at 10k per folder and asks
for under 100k files per repo; 9.3M loose files breaks both by two orders of magnitude.
The payload is a file tree, not tabular rows, so the dataset viewer does not apply.

## Groups

Video groups are named `video_data__<dataset>`, matching `video_data/<dataset>/` in the tree.

| Group | Shards | Size | What |
|---|---|---|---|
| `video_data__ViCaS` | 78 | 388.2 GiB | ViCaS grounded video captions + frames |
| `video_data__SA-FARI` | 37 | 242.2 GiB | SA-FARI wildlife video, frames + masklets |
| `Adobe_EntitySeg` | 38 | 193.9 GiB | EntitySeg images + annotations |
| `coco` | 25 | 143.2 GiB | COCO 2014/2017 images, annotations, caches |
| `video_data__BOSTVG` | 11 | 55.0 GiB | BOSTVG / OmniSTVG + SAM3 masks |
| `GoldG` | 12 | 53.0 GiB | GoldG (Flickr30k entities + GQA) |
| `rf100` | 9 | 43.4 GiB | RF100-VL |
| `coconut` | 10 | 41.7 GiB | COCONut panoptic PNGs + caches |
| `video_data__VIPOSeg` | 4 | 17.7 GiB | VIPOSeg frames + panomasks |
| `SA-1B` | 4 | 17.1 GiB | SA-1B annotation caches (not the images) |
| `pixno_points` | 3 | 9.9 GiB | PixMo pointing caches |
| `refclef` | 2 | 5.6 GiB | RefCLEF / saiapr_tc-12 |
| 22 more | 1 each | ~14 GiB | reasonseg, lvis, manual_annotation, refcoco/+/g, grefcoco, panoptic-narrative-grounding, PhraseCutDataset, pixmo, VidSTG, and the small video caches |

## Reconstructing the tree

```bash
hf download royguw/grounding-data --repo-type dataset --local-dir ./dl
GD=/your/path/to/grounding_data
for t in ./dl/shards/*/*.tar; do tar -xf "$t" -C "$GD"; done
```

To pull one dataset only, read its index and fetch just the shards it names:

```python
import pyarrow.parquet as pq
from huggingface_hub import hf_hub_download

idx = pq.read_table(hf_hub_download("royguw/grounding-data", "index/coco.parquet",
                                    repo_type="dataset")).to_pydict()
for shard in dict.fromkeys(idx["shard"]):
    hf_hub_download("royguw/grounding-data", shard, repo_type="dataset")
```

The index is also how you find a single file without downloading everything: look up its
path, and it tells you the one shard to fetch.

### A note on paths

The annotation parquets store **absolute** source paths in columns such as `image_path`,
`png_path` and `frames_dir`. After extracting you must rewrite those prefixes to wherever
you put the data. Two prefixes appear:

- paths under the grounding_data root -> `$GD`
- paths under `mm-olmo/` -> wherever you extracted
  [`royguw/mm-olmo-images`](https://huggingface.co/datasets/royguw/mm-olmo-images)

Video frame lists are stored as extension-less stems in `frame_basenames_json`, resolved
against the directory listing; where that column is absent, frames are `f"{i:05d}"` for
`i in range(num_frames)`.

## Licensing and attribution

This is a **redistribution mirror of third-party datasets** assembled for reproducibility.
It is not a new dataset, and no new license is claimed over the underlying images, video
frames or upstream annotations.

The groups carry different upstream licenses — several are research or non-commercial only
(EntitySeg, for one, is Adobe non-commercial), and some require accepting upstream terms
before use. **Check the license of any group you intend to use and comply with it.** If you
are an author or rights-holder of an upstream dataset here and want a subset removed, open
a discussion on this repo and it will be taken down.