Rewrite dataset card around data domains and usage

README.md

@@ -5,229 +5,270 @@ task_categories:
 tags:
   - visual-grounding
   - pointing
+  - document-layout-analysis
+  - gui-grounding
   - megatron-energon
   - webdataset
   - multimodal
-size_categories:
-  - 10M<n<100M
 ---
 
 # LocateAnything-Data
 
-[中文说明](README_CN.md) | [Hugging Face repository](https://huggingface.co/datasets/NVEagle/LocateAnything-Data)
-
-> Official release repository:
-> [`NVEagle/LocateAnything-Data`](https://huggingface.co/datasets/NVEagle/LocateAnything-Data).
-> This collection does not declare a single umbrella license. Before using any
-> component, consult and comply with the license and terms of its original
-> upstream dataset.
-
-LocateAnything-Data is a spatial vision-language training collection packaged
-for indexed JSONL access and Megatron-Energon. It contains object detection,
-visual grounding, and pointing records. It deliberately excludes VQA,
-benchmark, test, validation, chat, and internal build-provenance data.
-
-Dataset inventory:
-
-| Item | Count |
-|---|---:|
-| Dataset folders | 42 |
-| Annotation views | 47 |
-| Training records | 10,136,648 |
-| Canonical media pools | 41 |
-| Referenced physical images | 9,801,026 total; 7,402,667 hosted |
-| Canonical TAR shards | 926 in the source inventory; 583 hosted |
-
-OS-Atlas is part of this unified 42-dataset release. It is not published as a
-second copy or a separate media repository.
-
-## Media redistribution boundary
-
-All 47 annotation views and the sanitized record-to-source-image mappings are
-in release scope. Images from the following seven upstream sources are not
-redistributed. Users must obtain those images from the official source and
-hydrate them locally.
-
-| Upstream source | LocateAnything dataset ID | Canonical pool ID | Records | Official access | Expected source-relative path |
-|---|---|---|---:|---|---|
-| CrowdHuman | `crowdhuman` | `locany--crowdhuman--train--source_image--v000001` | 15,000 | [download](https://www.crowdhuman.org/download.html) | `train/Images/<name>.jpg` |
-| DeepFashion2 | `deepfashion2` | `locany--deepfashion2--train--source_image--v000001` | 191,961 | [instructions](https://github.com/switchablenorms/DeepFashion2), [request form](https://docs.google.com/forms/d/e/1FAIpQLSeIoGaFfCQILrtIZPykkr8q_h9qQ5BoTYbjvf95aXbid0v2Bw/viewform) | `train/image/000001.jpg` |
-| Flickr30K | `flickr30k` | `locany--flickr30k--train--source_image--v000001` | 29,781 | [official page](https://shannon.cs.illinois.edu/DenotationGraph/data/index.html) | `flickr30k-images/<flickr_id>.jpg` |
-| ImageNet | `partimagenet` | `locany--imagenet--train--source_image--v000001` | 100,000 | [download/request access](https://www.image-net.org/download.php) | `train/<wnid>/<wnid>_*.JPEG` |
-| Objects365 | `object365` | `locany--object365--train--source_image--v000001` | 1,742,289 | [download and terms](https://www.objects365.org/download.html) | `images/train/patch*/objects365_v1_*.jpg` |
-| SKU-110K | `sku110k` | `locany--sku110k--train--source_image--v000001` | 8,219 | [author repository](https://github.com/eg4000/SKU110K_CVPR19) | `images/train_*.jpg` |
-| Unsplash | `unsplash` | `locany--unsplash--train--source_image--v000001` | 329,721 | [dataset access](https://unsplash.com/data), [dataset terms](https://github.com/unsplash/datasets/blob/master/TERMS.md) | `raw_unsplash_images/<photo_id>.jpg` |
-
-ImageNet is a media source, not a LocateAnything dataset folder. Its images are
-used by the `partimagenet` annotation view. Objects365 is the official upstream
-name; `object365` is the stable LocateAnything dataset ID.
-
-These seven sources account for 2,416,971 records, 2,398,359 physical images,
-and 343 internal TAR shards. None of those TAR payloads are part of the public
-repository. The remaining hosted subset contains 35 dataset folders,
-7,719,677 records, 34 media pools, 7,402,667 physical images, and 583 TAR
-shards. Every hosted TAR has exactly one repository path.
-
-This repository is not a mirror of the seven upstream datasets and grants no
-rights to their media. Obtain the media independently, accept the terms in
-effect at download time, and preserve filenames and directory layout. The
-hydration tooling does not bypass registration, request forms, passwords, or
-access controls.
-
-Additional source-specific notes:
-
-- CrowdHuman requires the training image archives referenced by the mapping.
-  Its official page states that the images must not be redistributed.
-- DeepFashion2 requires a per-user access request and a password for the
-  archives. The released mapping references `train/image`.
-- Flickr30K images remain subject to Flickr's terms. The official project also
-  offers a publicly distributable links-and-captions version.
-- The ImageNet view uses a 100,000-image subset of the ILSVRC-style training
-  layout. Do not flatten synset directories.
-- Objects365 users should preserve the V1/V2 package filenames and use the
-  official annotation `file_name` values rather than renaming images.
-- SKU-110K users should preserve the extracted image and CSV layout.
-- The Unsplash Lite release is not necessarily sufficient for this 329,721
-  image mapping. Obtain the matching authorized release and require SHA-256
-  verification; do not substitute similarly named images.
-
-## Repository layout
-
-The repository keeps one canonical path for every payload:
+[中文](README_CN.md) · [Paper](https://arxiv.org/abs/2605.27365) ·
+[Model](https://huggingface.co/nvidia/LocateAnything-3B) ·
+[Code](https://github.com/NVlabs/Eagle/tree/main/Embodied)
+
+## Overview
+
+LocateAnything-Data is the public training-data release for
+**LocateAnything: Fast and High-Quality Vision-Language Grounding with Parallel
+Box Decoding**.
+
+LocateAnything formulates detection and visual grounding as a unified
+vision-language task. Given an image and a category, phrase, text string, or
+action-oriented instruction, the model predicts the corresponding bounding
+boxes or points. The data spans natural images, dense scenes, people,
+autonomous driving, embodied interaction, graphical user interfaces, scene
+text, documents, and tables.
+
+The paper's Parallel Box Decoding treats a box or point as a structured
+geometric unit instead of generating its coordinates independently.
+LocateAnything-Data provides the diverse spatial supervision used to train
+this unified formulation across visual domains.
+
+This repository provides:
+
+- detection, grounding, and pointing annotations in previewable JSONL;
+- image media packed as indexed WebDataset TAR shards;
+- Megatron-Energon metadata for distributed training; and
+- public mappings from every training record and packed image back to its
+  original dataset-relative media name.
+
+## Data coverage
+
+The following grouping describes the main visual domains in this release.
+Several datasets naturally support more than one domain or task.
+
+| Domain | Included datasets |
+|---|---|
+| General detection and grounding | COCO, RefCOCO, RefCOCO+, RefCOCOg, G-RefCOCO, Flickr30K, HumanRef-CoT, Objects365, Open Images V6, OWDOD, PACO, PartImageNet, Unsplash, V3Det |
+| Dense object detection | CrowdHuman, DeepFashion2, HumanParts, MOT17Det, MOT20Det, SKU-110K |
+| Autonomous driving and robotics | BDD100K, EgoObjects, nuImages, PixMo Points, RoboAfford |
+| GUI | GroundCUA, OS-Atlas, ScaleCUA |
+| Text recognition | ArT, BLIP3-OCR, IDL-OCR, HierText, ICDAR 2013, ICDAR 2015, LSVT, RCTW, ReCTS, SROIE, TextOCR, WildReceipt |
+| Documents, layouts, and tables | CDLA, DocLayNet, M6Doc, PubLayNet, TableBank, TabRecSet |
+
+### Spatial supervision
+
+All records use a shared `query` representation:
+
+| `task_type` | Query target | Typical use |
+|---|---|---|
+| `detection_grounding` | Bounding box `[x1, y1, x2, y2]` | Category detection, phrase grounding, GUI grounding, OCR, and layout localization |
+| `pointing` | Point `[x, y]` | Point-based localization |
+| `detection_grounding_pointing` | Boxes and points in one record | Affordance and mixed spatial supervision |
+
+Coordinates use the LocateAnything normalized spatial grid and can be mapped
+to image pixels with `x * width / 1000` and `y * height / 1000`.
+
+## Release format
+
+Each dataset has its own folder and may contain one or more annotation
+**views**. A view is a training annotation variant over a media source—for
+example, COCO detection and RefCOCO grounding are separate views while sharing
+the same underlying images.
 
 ```text
 LocateAnything-Data/
-├── .gitattributes
-├── README.md
-├── README_CN.md
-├── release-policy.json
-├── metadataset.yaml                 # immediately usable hosted subset
-├── metadataset-full.yaml            # all 42 datasets after local hydration
-├── manifests/
-│   ├── release.json
-│   ├── files.jsonl
-│   ├── datasets.jsonl
-│   ├── views.jsonl
-│   ├── media-pools.jsonl
-│   └── restricted-media.json
+├── metadataset.yaml
+├── metadataset-full.yaml
 ├── datasets/
 │   └── <dataset_id>/
 │       ├── metadataset.yaml
-│       └── views/
-│           └── <view_id>/
-│               ├── metadataset.yaml
-│               ├── records.jsonl
-│               └── records.jsonl.idx
+│       └── views/<view_id>/
+│           ├── metadataset.yaml
+│           ├── records.jsonl
+│           └── records.jsonl.idx
 ├── media/
 │   └── <pool_id>/
-│       ├── availability.json
-│       ├── .nv-meta/                # present only for hosted media
-│       └── shards/                  # present only for hosted media
+│       ├── .nv-meta/
+│       └── shards/
 │           ├── image-NNNNNNNN.tar
 │           └── image-NNNNNNNN.tar.idx
 ├── mappings/
-│   ├── media/<pool_id>/part-NNNNN.parquet
-│   └── views/<dataset_id>/<view_id>/part-NNNNN.parquet
-├── schemas/
-│   ├── media-source-map.schema.json
-│   └── view-source-map.schema.json
+│   ├── media/<pool_id>/part-*.parquet
+│   └── views/<dataset_id>/<view_id>/part-*.parquet
 ├── examples/
-│   ├── read_indexed_jsonl.py
-│   └── read_energon.py
 └── tools/
-    ├── hydrate_restricted_media.py
-    └── verify_release.py
 ```
 
-There is intentionally no `datasets/<dataset_id>/media/` tree. From
-`datasets/<dataset_id>/views/<view_id>/metadataset.yaml`, every auxiliary media
-path must be exactly `../../../../media/<pool_id>`. The seven external pools
-contain only `availability.json` until a user hydrates them locally.
+- `records.jsonl` keeps annotations easy to inspect and process.
+- `records.jsonl.idx` stores byte offsets for direct row access.
+- Uncompressed TAR shards preserve encoded image bytes and support indexed
+  reads without extracting millions of files.
+- `.tar.idx` and `.nv-meta` are consumed by Megatron-Energon.
+- Parquet mappings preserve record lineage and original media names.
 
-`metadataset.yaml` blends the 35 datasets whose media is hosted on Hugging
-Face. `metadataset-full.yaml` blends all 42 datasets and must fail fast until
-the seven external pools have been hydrated and validated.
+### Annotation record
 
-## Annotation format
-
-Annotations remain human-previewable JSON Lines. One line is one independent
-training record. The following is an abridged schema example:
+Each JSONL line is one independent training record:
 
 ```json
 {
   "_source": {
-    "schema": "eagle-energon.source-record/v1",
-    "dataset_id": "crowdhuman",
-    "view_id": "locate_anything_crowdhuman_crowdhuman",
-    "sample_id": "crowdhuman:locate_anything_crowdhuman_crowdhuman:file:0009:row:000000000000",
-    "source_annotation": "annotations/CrowdHuman/CrowdHuman.jsonl",
-    "source_file_index": 9,
-    "source_row": 0,
-    "source_record_sha256": "<sha256>"
+    "dataset_id": "coco",
+    "view_id": "locate_anything_coco_refcoco",
+    "sample_id": "coco:locate_anything_coco_refcoco:..."
   },
   "image": {
     "kind": "image",
     "source": "media_00",
-    "path": "m/000000000082.jpg"
+    "path": "m/000000012345.jpg"
   },
   "query": {
-    "Person": [[1, 410, 168, 1000]]
+    "the person holding an umbrella": [[124, 86, 612, 944]]
   },
   "task_type": "detection_grounding"
 }
 ```
 
-`query` values are numeric points `[x, y]` or boxes
-`[x1, y1, x2, y2]`. The reader does not rescale or rewrite them.
-`image.source` selects the view's Energon auxiliary pool and `image.path` is
-the exact TAR member name. Never infer a source filename from the packed
-member name.
+`image.source` selects the auxiliary media pool declared by the view-level
+`metadataset.yaml`. `image.path` is the exact member name inside the TAR.
 
-`records.jsonl.idx` is a little-endian uint64 array with `N + 1` byte offsets
-for `N` records. Row `i` occupies
-`[offset[i], offset[i + 1])`, enabling O(1) indexed record access.
+## Download
 
-Preview records:
+Install the Hugging Face CLI and download the repository:
+
+```bash
+pip install -U huggingface_hub
+
+hf download NVEagle/LocateAnything-Data \
+  --repo-type dataset \
+  --local-dir LocateAnything-Data
+
+cd LocateAnything-Data
+```
+
+Use `hf download ... --dry-run` or `--include` when you only need selected
+datasets.
+
+## Inspect annotations
+
+JSONL records can be previewed with standard command-line tools:
 
 ```bash
 sed -n '1,3p' \
-  datasets/object365/views/locate_anything_object365_object365/records.jsonl \
-  | jq -c .
+  datasets/coco/views/locate_anything_coco_refcoco/records.jsonl \
+  | jq .
 ```
 
-Read one indexed row without loading the file:
+Read a row by index without scanning the JSONL file:
 
 ```bash
 python examples/read_indexed_jsonl.py \
-  datasets/object365/views/locate_anything_object365_object365/records.jsonl \
+  datasets/coco/views/locate_anything_coco_refcoco/records.jsonl \
   100
 ```
 
-## Public source-image mapping
+## Visualize a sample
 
-The public repository does not contain the internal SQLite mapping databases,
-which include absolute paths and filesystem evidence. Instead, it contains two
-normalized, Zstandard-compressed Parquet relations:
+Install the runtime dependencies:
 
-1. `mappings/media/<pool_id>/...` has one row per upstream source-image alias.
-   It maps the original dataset-relative filename/path and SHA-256 to a
-   canonical `member_name` and content SHA-256.
-2. `mappings/views/<dataset_id>/<view_id>/...` has one row per record-media
-   occurrence. It maps the record lineage and JSON pointer to the exact
-   pool-local `source_id` and canonical member.
+```bash
+pip install "megatron-energon==7.4.0" pillow
+```
 
-The second relation is required because byte-identical images can have multiple
-upstream filenames. Joining only on `member_name` would be ambiguous. The
-authoritative join is `(pool_id, source_id)`.
-`source_snapshot_id` identifies the conversion/source-manifest snapshot; it
-must not be interpreted as an upstream dataset release name.
+The following example loads one sample through Energon and draws all boxes and
+points:
+
+```python
+from io import BytesIO
+from pathlib import Path
+
+from PIL import Image, ImageDraw
+from megatron.energon import (
+    WorkerConfig,
+    get_savable_loader,
+    get_train_dataset,
+)
+
+from examples.read_energon import LocateAnythingTaskEncoder
+
+dataset = get_train_dataset(
+    Path("datasets/ocr_icdar2013/metadataset.yaml"),
+    split_part="train",
+    worker_config=WorkerConfig(
+        rank=0, world_size=1, num_workers=0
+    ),
+    batch_size=None,
+    max_samples_per_sequence=1,
+    shuffle_over_epochs_multiplier=1,
+    shuffle_buffer_size=None,
+    task_encoder=LocateAnythingTaskEncoder(),
+    repeat=False,
+)
+sample = next(iter(get_savable_loader(dataset)))
+
+image = Image.open(BytesIO(sample.image)).convert("RGB")
+draw = ImageDraw.Draw(image)
+
+def to_pixel(target):
+    values = list(target)
+    for i in range(0, len(values), 2):
+        values[i] = round(values[i] * image.width / 1000)
+        values[i + 1] = round(values[i + 1] * image.height / 1000)
+    return values
+
+for label, targets in sample.record["query"].items():
+    for target in targets:
+        xy = to_pixel(target)
+        if len(xy) == 4:
+            draw.rectangle(xy, outline="red", width=3)
+            draw.text((xy[0], xy[1]), label, fill="red")
+        elif len(xy) == 2:
+            x, y = xy
+            draw.ellipse((x - 5, y - 5, x + 5, y + 5), fill="red")
+            draw.text((x + 6, y), label, fill="red")
+
+image.save("sample.png")
+```
 
-Example: recover the exact original source image for record row 12,911:
+## Read for training
 
-```sql
-INSTALL parquet;
-LOAD parquet;
+The repository targets `megatron-energon==7.4.0`. Choose the entry point that
+matches your training scope:
+
+| Entry point | Purpose |
+|---|---|
+| `metadataset.yaml` | All datasets whose media is available directly in this repository |
+| `metadataset-full.yaml` | The full mixture after locally adding the externally downloaded media |
+| `datasets/<dataset_id>/metadataset.yaml` | One dataset |
+| `datasets/<dataset_id>/views/<view_id>/metadataset.yaml` | One annotation view |
+
+Run the included reader:
 
+```bash
+python examples/read_energon.py \
+  datasets/coco/metadataset.yaml \
+  --workers 8 \
+  --samples 16
+```
+
+`examples/read_energon.py` contains a minimal task encoder. It resolves the
+record's media source, reads the encoded image directly from the indexed TAR,
+and returns the annotation dictionary with image bytes.
+
+## Map a record back to its source image
+
+Two Parquet relations connect each training record to its source image:
+
+- `mappings/views/...` connects a JSONL row to a media `source_id`.
+- `mappings/media/...` connects that `source_id` to the original
+  dataset-relative filename, file hash, and packed TAR member.
+
+The authoritative join key is `(pool_id, source_id)`. For example:
+
+```sql
 SELECT
   v.row_id,
   v.source_sample_id,
@@ -242,59 +283,26 @@ JOIN read_parquet(
   'mappings/media/locany--crowdhuman--train--source_image--v000001/*.parquet'
 ) AS m
 USING (pool_id, source_id)
-WHERE v.row_id = 12911;
-```
-
-Reverse lookup from an upstream path to all LocateAnything records:
-
-```sql
-SELECT v.dataset_id, v.view_id, v.row_id, v.source_sample_id
-FROM read_parquet(
-  'mappings/media/locany--crowdhuman--train--source_image--v000001/*.parquet'
-) AS m
-JOIN read_parquet(
-  'mappings/views/crowdhuman/locate_anything_crowdhuman_crowdhuman/*.parquet'
-) AS v
-USING (pool_id, source_id)
-WHERE m.source_relative_path =
-  'train/Images/273271,1017c000ac1360b7.jpg';
+WHERE v.row_id = 100;
 ```
 
-The public mapping preserves original source filenames, not LMDB keys or
-packed member names. It never contains `source_absolute_path`, internal
-annotation paths, Lustre paths, usernames, inode/device metadata, symlink
-targets, access tokens, cookies, or signed URLs. See the schemas under
-`schemas/`.
-
-## Reading with Megatron-Energon
+## Media downloaded from upstream sources
 
-The storage contract targets `megatron-energon==7.4.0`. Use:
+Media from the following datasets must be obtained from the original source.
+Their annotations and source-image mappings are included here.
 
-- `metadataset.yaml` for the immediately readable hosted subset;
-- `metadataset-full.yaml` after all seven external pools are hydrated;
-- `datasets/<dataset_id>/metadataset.yaml` for one dataset; or
-- a view-level `metadataset.yaml` for debugging.
+| Dataset folder | Media source |
+|---|---|
+| `crowdhuman` | [CrowdHuman](https://www.crowdhuman.org/download.html) |
+| `deepfashion2` | [DeepFashion2](https://github.com/switchablenorms/DeepFashion2) |
+| `flickr30k` | [Flickr30K](https://shannon.cs.illinois.edu/DenotationGraph/data/index.html) |
+| `partimagenet` | [ImageNet](https://www.image-net.org/download.php) |
+| `object365` | [Objects365](https://www.objects365.org/download.html) |
+| `sku110k` | [SKU-110K](https://github.com/eg4000/SKU110K_CVPR19) |
+| `unsplash` | [Unsplash Dataset](https://unsplash.com/data) |
 
-The format supports random indexed JSONL rows, indexed uncompressed TAR
-samples, multiple workers, rank-aware partitioning, and stream packing. For an
-exact without-replacement sample permutation, use one indexed row per
-Energon slice (`max_samples_per_sequence=1`), one epoch shuffle multiplier,
-and no second shuffle buffer. Packing should happen after sample
-randomization.
-
-The repository provides a small task encoder and runnable loader in
-`examples/read_energon.py`:
-
-```bash
-python examples/read_energon.py metadataset.yaml --workers 8 --samples 3
-```
-
-A missing external pool is an error: training code must not silently skip
-records.
-
-## Hydrating the seven external media pools
-
-The hydrator uses the public mapping as the sole file list:
+After downloading a source dataset with its original directory and filenames,
+build the local media shards from the public mapping:
 
 ```bash
 python tools/hydrate_restricted_media.py \
@@ -303,44 +311,16 @@ python tools/hydrate_restricted_media.py \
   --source-root /data/upstream/CrowdHuman
 ```
 
-For every mapped source image it:
-
-1. resolve the exact `source_relative_path`, never basename-guess;
-2. verify byte length and SHA-256;
-3. preserve the stable `member_name`;
-4. content-deduplicate within the canonical pool;
-5. write approximately 4 GiB uncompressed TAR shards;
-6. build `.tar.idx` and official Energon `.nv-meta`;
-7. verify every annotation reference; and
-8. write a local receipt.
-
-Hydrated media remains local and must never be pushed back to the Hugging Face
-repository.
-
-## Canonical storage and deduplication
-
-- `NVEagle/LocateAnything-Data` is the collection's single canonical Hugging
-  Face dataset repository.
-- Every canonical TAR has exactly one
-  `media/<pool_id>/shards/` repository path.
-- Dataset-local media projections are not part of the repository.
-- OS-Atlas belongs to the unified catalog and has no separate copy.
-- `manifests/files.jsonl` lists every release object except itself, with one
-  row per canonical repository path.
-
-Hugging Face Xet may deduplicate chunks internally; this does not change the
-path-level storage contract above.
-
-## Licenses and upstream terms
-
-LocateAnything-Data combines annotations and references derived from multiple
-upstream datasets, so this repository does not select or imply one unified
-license for all components. Each upstream dataset remains governed by its own
-license, terms of use, access conditions, and attribution requirements. Users
-must review the official upstream source linked in this README and comply with
-the applicable terms before downloading, hydrating, using, or redistributing
-any component. This repository does not grant additional rights to upstream
-media.
+The tool verifies mapped files and creates the TAR indexes and Energon metadata
+expected by `metadataset-full.yaml`.
+
+## License
+
+LocateAnything-Data is assembled from multiple upstream datasets. Each
+component remains subject to the license, terms of use, access requirements,
+and attribution requirements of its original source. This repository does not
+grant additional rights to upstream images. Review the corresponding upstream
+terms before downloading, using, or redistributing a component.
 
 ## Citation
 

README_CN.md

@@ -5,220 +5,259 @@ task_categories:
 tags:
   - visual-grounding
   - pointing
+  - document-layout-analysis
+  - gui-grounding
   - megatron-energon
   - webdataset
   - multimodal
-size_categories:
-  - 10M<n<100M
 ---
 
 # LocateAnything-Data
 
-[English](README.md) | [Hugging Face 仓库](https://huggingface.co/datasets/NVEagle/LocateAnything-Data)
-
-> 官方发布仓库：
-> [`NVEagle/LocateAnything-Data`](https://huggingface.co/datasets/NVEagle/LocateAnything-Data)。
-> 本集合不声明一个覆盖全部内容的统一许可证。使用任何组成部分前，请查阅并
-> 遵守对应原始上游数据集的许可证和使用条款。
-
-LocateAnything-Data 是面向空间视觉语言训练的数据集合，使用可索引 JSONL 和
-Megatron-Energon 组织。任务只包括目标检测、visual grounding 和 pointing；
-不包含 VQA、benchmark、test、validation、chat 或内部构建 provenance。
-
-数据集总盘点如下：
-
-| 项目 | 数量 |
-|---|---:|
-| Dataset folder | 42 |
-| 标注 view | 47 |
-| 训练 record | 10,136,648 |
-| Canonical media pool | 41 |
-| 被引用的物理图片 | 总计 9,801,026；其中托管 7,402,667 |
-| Canonical TAR shard | 源盘点 926；其中托管 583 |
-
-OS-Atlas 已经属于这个统一的 42-dataset 发布，不再单独建立一份媒体仓库或重复
-发布。
-
-## 图片再分发边界
-
-47 个标注 view 和脱敏后的 record-to-source-image mapping 都属于开源范围。
-以下 7 个上游来源的图片不能随仓库再分发，使用者必须从官方入口自行取得并在
-本地 hydrate：
-
-| 上游来源 | LocateAnything dataset ID | Canonical pool ID | Record | 官方入口 | 预期 source-relative path |
-|---|---|---|---:|---|---|
-| CrowdHuman | `crowdhuman` | `locany--crowdhuman--train--source_image--v000001` | 15,000 | [下载](https://www.crowdhuman.org/download.html) | `train/Images/<name>.jpg` |
-| DeepFashion2 | `deepfashion2` | `locany--deepfashion2--train--source_image--v000001` | 191,961 | [官方说明](https://github.com/switchablenorms/DeepFashion2)、[申请表](https://docs.google.com/forms/d/e/1FAIpQLSeIoGaFfCQILrtIZPykkr8q_h9qQ5BoTYbjvf95aXbid0v2Bw/viewform) | `train/image/000001.jpg` |
-| Flickr30K | `flickr30k` | `locany--flickr30k--train--source_image--v000001` | 29,781 | [官方页面](https://shannon.cs.illinois.edu/DenotationGraph/data/index.html) | `flickr30k-images/<flickr_id>.jpg` |
-| ImageNet | `partimagenet` | `locany--imagenet--train--source_image--v000001` | 100,000 | [下载/申请访问](https://www.image-net.org/download.php) | `train/<wnid>/<wnid>_*.JPEG` |
-| Objects365 | `object365` | `locany--object365--train--source_image--v000001` | 1,742,289 | [下载及条款](https://www.objects365.org/download.html) | `images/train/patch*/objects365_v1_*.jpg` |
-| SKU-110K | `sku110k` | `locany--sku110k--train--source_image--v000001` | 8,219 | [作者仓库](https://github.com/eg4000/SKU110K_CVPR19) | `images/train_*.jpg` |
-| Unsplash | `unsplash` | `locany--unsplash--train--source_image--v000001` | 329,721 | [数据入口](https://unsplash.com/data)、[Dataset Terms](https://github.com/unsplash/datasets/blob/master/TERMS.md) | `raw_unsplash_images/<photo_id>.jpg` |
-
-ImageNet 在这里是媒体来源，不是一个 LocateAnything dataset folder；使用它的标注
-数据集是 `partimagenet`。Objects365 是上游官方名称，`object365` 只是
-LocateAnything 中的稳定 ID。
-
-这 7 个来源覆盖 2,416,971 条 record、2,398,359 个物理图片和内部 343 个 TAR；
-这些 TAR payload 一个都不会进入公开仓库。可以直接托管的其余部分包含 35 个
-dataset folder、7,719,677 条 record、34 个媒体池、7,402,667 个物理图片和
-583 个 TAR。每个托管 TAR 在仓库里只有一个路径。
-
-本仓库不是上述 7 个上游数据集的镜像，也不授予其媒体权利。使用者必须自行从
-官方入口取得媒体、接受下载时有效的条款，并保持原始文件名和目录结构。
-Hydration 工具不会绕过注册、申请表、密码或访问控制。
-
-补充下载说明：
-
-- CrowdHuman 需要下载 mapping 所引用的训练图片分卷；官网明确规定图片不得
-  再分发。
-- DeepFashion2 需要逐用户申请并取得压缩包密码；公开 mapping 引用
-  `train/image`。
-- Flickr30K 图片继续受 Flickr 条款约束；官方项目另行提供可公开分发的图片链接
-  与 captions 版本。
-- ImageNet view 使用 ILSVRC 风格训练目录中的 100,000 张图片，不能把 synset
-  目录打平。
-- Objects365 应保留 V1/V2 包的文件名，并按官方标注中的 `file_name` 关联，
-  不能自行改名。
-- SKU-110K 应保持解压后的图片和 CSV 目录结构。
-- Unsplash Lite 不一定覆盖本 mapping 的 329,721 张图片。必须取得匹配的授权
-  release 并用 SHA-256 校验，不能用同名或相似图片替换。
-
-## 仓库组织
-
-仓库对每个 payload 只保留一个 canonical 路径：
+[English](README.md) · [论文](https://arxiv.org/abs/2605.27365) ·
+[模型](https://huggingface.co/nvidia/LocateAnything-3B) ·
+[代码](https://github.com/NVlabs/Eagle/tree/main/Embodied)
+
+## 简介
+
+LocateAnything-Data 是
+**LocateAnything: Fast and High-Quality Vision-Language Grounding with
+Parallel Box Decoding** 的公开训练数据。
+
+LocateAnything 将目标检测和视觉定位统一为视觉语言任务：给定图片，以及类别、
+描述短语、文本内容或动作指令，模型预测相应的边界框或点。数据覆盖自然图像、
+密集场景、人物、自动驾驶、具身交互、图形界面、场景文字、文档和表格。
+
+论文提出的 Parallel Box Decoding 将一个框或点作为完整的几何单元，而不是分别
+生成各个坐标。LocateAnything-Data 为这一统一任务提供跨视觉领域的多样化空间
+监督。
+
+本仓库提供：
+
+- 使用 JSONL 保存、方便预览的检测、grounding 和 pointing 标注；
+- 使用带索引的 WebDataset TAR 保存的图片；
+- 用于分布式训练的 Megatron-Energon 元数据；
+- 从每条训练标注和打包图片回溯到原始数据集相对路径的公开 mapping。
+
+## 数据范围
+
+下表按照主要视觉领域组织数据。部分数据集天然覆盖多个领域或任务。
+
+| 领域 | 包含的数据集 |
+|---|---|
+| 通用检测与 Grounding | COCO、RefCOCO、RefCOCO+、RefCOCOg、G-RefCOCO、Flickr30K、HumanRef-CoT、Objects365、Open Images V6、OWDOD、PACO、PartImageNet、Unsplash、V3Det |
+| 密集目标检测 | CrowdHuman、DeepFashion2、HumanParts、MOT17Det、MOT20Det、SKU-110K |
+| 自动驾驶和机器人 | BDD100K、EgoObjects、nuImages、PixMo Points、RoboAfford |
+| GUI | GroundCUA、OS-Atlas、ScaleCUA |
+| 文字识别 | ArT、BLIP3-OCR、IDL-OCR、HierText、ICDAR 2013、ICDAR 2015、LSVT、RCTW、ReCTS、SROIE、TextOCR、WildReceipt |
+| 文档、版面与表格 | CDLA、DocLayNet、M6Doc、PubLayNet、TableBank、TabRecSet |
+
+### 空间监督类型
+
+所有标注使用统一的 `query` 表达：
+
+| `task_type` | 标注目标 | 典型用途 |
+|---|---|---|
+| `detection_grounding` | 边界框 `[x1, y1, x2, y2]` | 类别检测、短语定位、GUI 定位、OCR 和版面定位 |
+| `pointing` | 点 `[x, y]` | 基于点的目标定位 |
+| `detection_grounding_pointing` | 同一条标注中的框和点 | Affordance 和混合空间监督 |
+
+坐标使用 LocateAnything 的归一化空间网格。映射到图片像素时使用
+`x * width / 1000` 和 `y * height / 1000`。
+
+## 开源格式
+
+每个数据集都有独立目录，并且可以包含一个或多个标注 **view**。View 表示同一
+媒体源上的一种训练标注，例如 COCO detection 和 RefCOCO grounding 是不同
+view，但可以共享底层图片。
 
 ```text
 LocateAnything-Data/
-├── .gitattributes
-├── README.md
-├── README_CN.md
-├── release-policy.json
-├── metadataset.yaml                 # 可立即读取的托管子集
-├── metadataset-full.yaml            # 本地补齐 7 个媒体池后的全集
-├── manifests/
-│   ├── release.json
-│   ├── files.jsonl
-│   ├── datasets.jsonl
-│   ├── views.jsonl
-│   ├── media-pools.jsonl
-│   └── restricted-media.json
+├── metadataset.yaml
+├── metadataset-full.yaml
 ├── datasets/
 │   └── <dataset_id>/
 │       ├── metadataset.yaml
-│       └── views/
-│           └── <view_id>/
-│               ├── metadataset.yaml
-│               ├── records.jsonl
-│               └── records.jsonl.idx
+│       └── views/<view_id>/
+│           ├── metadataset.yaml
+│           ├── records.jsonl
+│           └── records.jsonl.idx
 ├── media/
 │   └── <pool_id>/
-│       ├── availability.json
-│       ├── .nv-meta/                # 只存在于托管媒体池
-│       └── shards/                  # 只存在于托管媒体池
+│       ├── .nv-meta/
+│       └── shards/
 │           ├── image-NNNNNNNN.tar
 │           └── image-NNNNNNNN.tar.idx
 ├── mappings/
-│   ├── media/<pool_id>/part-NNNNN.parquet
-│   └── views/<dataset_id>/<view_id>/part-NNNNN.parquet
-├── schemas/
-│   ├── media-source-map.schema.json
-│   └── view-source-map.schema.json
+│   ├── media/<pool_id>/part-*.parquet
+│   └── views/<dataset_id>/<view_id>/part-*.parquet
 ├── examples/
-│   ├── read_indexed_jsonl.py
-│   └── read_energon.py
 └── tools/
-    ├── hydrate_restricted_media.py
-    └── verify_release.py
 ```
 
-仓库中禁止出现 `datasets/<dataset_id>/media/`。从
-`datasets/<dataset_id>/views/<view_id>/metadataset.yaml` 出发，每个 auxiliary
-媒体路径必须精确写成 `../../../../media/<pool_id>`。7 个外部媒体池在使用者
-本地 hydrate 前只有 `availability.json`。
-
-`metadataset.yaml` 混合 35 个图片已由 Hugging Face 托管的数据集；
-`metadataset-full.yaml` 混合全部 42 个数据集，在 7 个外部池全部 hydrate 并
-通过校验前必须 fail fast。
+- `records.jsonl` 便于直接查看和处理标注。
+- `records.jsonl.idx` 保存字节偏移，可以直接定位任意一行。
+- 未压缩 TAR 保留图片的原始编码字节，可以索引读取，无需解压成海量小文件。
+- `.tar.idx` 和 `.nv-meta` 供 Megatron-Energon 使用。
+- Parquet mapping 保存标注来源和原始媒体文件名。
 
-## 标注格式
+### 标注格式
 
-标注保持为方便预览的 JSON Lines，每行是一条独立训练 record。下面是删减后的
-schema 示例：
+JSONL 中的每一行是一条独立训练标注：
 
 ```json
 {
   "_source": {
-    "schema": "eagle-energon.source-record/v1",
-    "dataset_id": "crowdhuman",
-    "view_id": "locate_anything_crowdhuman_crowdhuman",
-    "sample_id": "crowdhuman:locate_anything_crowdhuman_crowdhuman:file:0009:row:000000000000",
-    "source_annotation": "annotations/CrowdHuman/CrowdHuman.jsonl",
-    "source_file_index": 9,
-    "source_row": 0,
-    "source_record_sha256": "<sha256>"
+    "dataset_id": "coco",
+    "view_id": "locate_anything_coco_refcoco",
+    "sample_id": "coco:locate_anything_coco_refcoco:..."
   },
   "image": {
     "kind": "image",
     "source": "media_00",
-    "path": "m/000000000082.jpg"
+    "path": "m/000000012345.jpg"
   },
   "query": {
-    "Person": [[1, 410, 168, 1000]]
+    "the person holding an umbrella": [[124, 86, 612, 944]]
   },
   "task_type": "detection_grounding"
 }
 ```
 
-`query` 中的值是数值点 `[x, y]` 或数值框 `[x1, y1, x2, y2]`，读取器不会
-再次缩放或改写。`image.source` 选择 view 的 Energon auxiliary pool，
-`image.path` 是 TAR 中的精确 member name。禁止根据 packed member name 猜测
-原始文件名。
+`image.source` 选择 view 级 `metadataset.yaml` 声明的辅助媒体池，
+`image.path` 是 TAR 内部的精确 member name。
 
-`records.jsonl.idx` 是 little-endian uint64 数组。N 条 record 对应 N+1 个
-uint64 offset（每个 8 字节），第 i 行范围为
-`[offset[i], offset[i + 1])`，因此可以 O(1) 定位。
+## 下载
 
-预览：
+安装 Hugging Face CLI 并下载仓库：
+
+```bash
+pip install -U huggingface_hub
+
+hf download NVEagle/LocateAnything-Data \
+  --repo-type dataset \
+  --local-dir LocateAnything-Data
+
+cd LocateAnything-Data
+```
+
+如果只需要部分数据集，可以先使用 `hf download ... --dry-run` 查看内容，并用
+`--include` 选择路径。
+
+## 查看标注
+
+JSONL 可以直接使用常见命令行工具预览：
 
 ```bash
 sed -n '1,3p' \
-  datasets/object365/views/locate_anything_object365_object365/records.jsonl \
-  | jq -c .
+  datasets/coco/views/locate_anything_coco_refcoco/records.jsonl \
+  | jq .
 ```
 
-不扫描整个文件、直接读取指定行：
+通过索引读取指定行，不需要扫描 JSONL：
 
 ```bash
 python examples/read_indexed_jsonl.py \
-  datasets/object365/views/locate_anything_object365_object365/records.jsonl \
+  datasets/coco/views/locate_anything_coco_refcoco/records.jsonl \
   100
 ```
 
-## 公开 source-image mapping
+## 可视化样本
 
-公开仓库不包含内部 SQLite mapping，因为其中含有绝对路径和文件系统证据。
-公开版使用两张规范化、Zstandard 压缩的 Parquet 关系表：
+安装运行依赖：
 
-1. `mappings/media/<pool_id>/...`：每个上游 source-image alias 一行，把原始
-   dataset-relative 文件名/路径及 SHA-256 映射到 canonical `member_name` 和
-   content SHA-256。
-2. `mappings/views/<dataset_id>/<view_id>/...`：每个 record-media occurrence
-   一行，把 record lineage 和 JSON pointer 映射到精确的 pool-local
-   `source_id` 及 canonical member。
+```bash
+pip install "megatron-energon==7.4.0" pillow
+```
 
-第二张表不可省略，因为字节完全相同的图片可能有多个上游文件名。只根据
-`member_name` 反查会产生歧义；唯一权威 join key 是
-`(pool_id, source_id)`。
-`source_snapshot_id` 表示转换/source-manifest snapshot，不能把它解释为
-上游数据集的 release 名称。
+下面的示例通过 Energon 读取一个样本，并绘制其中的所有框和点：
+
+```python
+from io import BytesIO
+from pathlib import Path
+
+from PIL import Image, ImageDraw
+from megatron.energon import (
+    WorkerConfig,
+    get_savable_loader,
+    get_train_dataset,
+)
+
+from examples.read_energon import LocateAnythingTaskEncoder
+
+dataset = get_train_dataset(
+    Path("datasets/ocr_icdar2013/metadataset.yaml"),
+    split_part="train",
+    worker_config=WorkerConfig(
+        rank=0, world_size=1, num_workers=0
+    ),
+    batch_size=None,
+    max_samples_per_sequence=1,
+    shuffle_over_epochs_multiplier=1,
+    shuffle_buffer_size=None,
+    task_encoder=LocateAnythingTaskEncoder(),
+    repeat=False,
+)
+sample = next(iter(get_savable_loader(dataset)))
+
+image = Image.open(BytesIO(sample.image)).convert("RGB")
+draw = ImageDraw.Draw(image)
+
+def to_pixel(target):
+    values = list(target)
+    for i in range(0, len(values), 2):
+        values[i] = round(values[i] * image.width / 1000)
+        values[i + 1] = round(values[i + 1] * image.height / 1000)
+    return values
+
+for label, targets in sample.record["query"].items():
+    for target in targets:
+        xy = to_pixel(target)
+        if len(xy) == 4:
+            draw.rectangle(xy, outline="red", width=3)
+            draw.text((xy[0], xy[1]), label, fill="red")
+        elif len(xy) == 2:
+            x, y = xy
+            draw.ellipse((x - 5, y - 5, x + 5, y + 5), fill="red")
+            draw.text((x + 6, y), label, fill="red")
+
+image.save("sample.png")
+```
 
-从第 12,911 行 record 找回精确上游图片：
+## 训练读取
 
-```sql
-INSTALL parquet;
-LOAD parquet;
+本仓库使用 `megatron-energon==7.4.0`。根据训练范围选择入口：
 
+| 入口 | 用途 |
+|---|---|
+| `metadataset.yaml` | 直接包含在本仓库中的全部媒体和标注 |
+| `metadataset-full.yaml` | 在本地补齐需要从上游下载的媒体后使用全集 |
+| `datasets/<dataset_id>/metadataset.yaml` | 读取单个数据集 |
+| `datasets/<dataset_id>/views/<view_id>/metadataset.yaml` | 读取单个标注 view |
+
+运行仓库提供的 reader：
+
+```bash
+python examples/read_energon.py \
+  datasets/coco/metadataset.yaml \
+  --workers 8 \
+  --samples 16
+```
+
+`examples/read_energon.py` 包含一个最小 task encoder。它根据标注解析媒体源，
+从带索引的 TAR 中直接读取编码图片，并返回标注字典和图片字节。
+
+## 从训练标注找回原始图片
+
+两组 Parquet 关系把每条训练标注关联到对应的原始图片：
+
+- `mappings/views/...`：把 JSONL 行关联到媒体 `source_id`。
+- `mappings/media/...`：把 `source_id` 关联到原始数据集相对路径、文件哈希和
+  TAR member。
+
+权威 join key 是 `(pool_id, source_id)`。例如：
+
+```sql
 SELECT
   v.row_id,
   v.source_sample_id,
@@ -233,54 +272,26 @@ JOIN read_parquet(
   'mappings/media/locany--crowdhuman--train--source_image--v000001/*.parquet'
 ) AS m
 USING (pool_id, source_id)
-WHERE v.row_id = 12911;
+WHERE v.row_id = 100;
 ```
 
-从原始路径反查所有 LocateAnything record：
+## 需要从上游下载的媒体
 
-```sql
-SELECT v.dataset_id, v.view_id, v.row_id, v.source_sample_id
-FROM read_parquet(
-  'mappings/media/locany--crowdhuman--train--source_image--v000001/*.parquet'
-) AS m
-JOIN read_parquet(
-  'mappings/views/crowdhuman/locate_anything_crowdhuman_crowdhuman/*.parquet'
-) AS v
-USING (pool_id, source_id)
-WHERE m.source_relative_path =
-  'train/Images/273271,1017c000ac1360b7.jpg';
-```
+以下数据集的媒体需要从原始来源获取；对应的标注和 source-image mapping 已经
+包含在本仓库中。
 
-公开 mapping 保留原始 source 文件名，不使用 LMDB key 或 packed member 冒充
-原始名字。mapping 绝不包含 `source_absolute_path`、内部 annotation 路径、
-Lustre 路径、用户名、inode/device、软链接目标、token、cookie 或 signed URL。
-精确字段见 `schemas/`。
+| 数据集目录 | 媒体来源 |
+|---|---|
+| `crowdhuman` | [CrowdHuman](https://www.crowdhuman.org/download.html) |
+| `deepfashion2` | [DeepFashion2](https://github.com/switchablenorms/DeepFashion2) |
+| `flickr30k` | [Flickr30K](https://shannon.cs.illinois.edu/DenotationGraph/data/index.html) |
+| `partimagenet` | [ImageNet](https://www.image-net.org/download.php) |
+| `object365` | [Objects365](https://www.objects365.org/download.html) |
+| `sku110k` | [SKU-110K](https://github.com/eg4000/SKU110K_CVPR19) |
+| `unsplash` | [Unsplash Dataset](https://unsplash.com/data) |
 
-## 使用 Megatron-Energon 读取
-
-格式绑定 `megatron-energon==7.4.0`：
-
-- 直接可读的托管子集使用 `metadataset.yaml`；
-- 7 个外部池全部 hydrate 后使用 `metadataset-full.yaml`；
-- 单数据集使用 `datasets/<dataset_id>/metadataset.yaml`；
-- 调试单 view 时使用对应 view 的 `metadataset.yaml`。
-
-该结构支持 JSONL 随机索引、未压缩 TAR 随机索引、多 worker、按 rank 切分和
-stream packing。完全无放回的 sample-level permutation 应使用
-`max_samples_per_sequence=1`、单 epoch shuffle multiplier，并关闭第二层
-shuffle buffer；packing 放在样本随机化之后执行。
-
-仓库在 `examples/read_energon.py` 中提供轻量 task encoder 和可运行的 loader：
-
-```bash
-python examples/read_energon.py metadataset.yaml --workers 8 --samples 3
-```
-
-外部媒体池缺失必须报错，训练代码不得静默跳过 record。
-
-## 本地补齐 7 个外部媒体池
-
-Hydrator 只以公开 mapping 作为文件清单：
+下载上游数据并保留原始目录和文件名后，可以根据公开 mapping 构建本地媒体
+shard：
 
 ```bash
 python tools/hydrate_restricted_media.py \
@@ -289,40 +300,14 @@ python tools/hydrate_restricted_media.py \
   --source-root /data/upstream/CrowdHuman
 ```
 
-它会对每个 source image：
-
-1. 精确解析 `source_relative_path`，禁止只按 basename 猜测；
-2. 校验字节数和 SHA-256；
-3. 保持稳定的 `member_name`；
-4. 在 canonical pool 内按内容去重；
-5. 写约 4 GiB 的未压缩 TAR；
-6. 生成 `.tar.idx` 和官方 Energon `.nv-meta`；
-7. 验证全部 annotation reference；
-8. 写本地 hydration receipt。
-
-Hydrate 生成的媒体只能保留在本地，绝不能回传 Hugging Face。
-
-## Canonical 存储与去重
-
-- `NVEagle/LocateAnything-Data` 是本集合唯一的 canonical Hugging Face
-  dataset repository。
-- 每个 canonical TAR 在仓库中只有一个
-  `media/<pool_id>/shards/` 路径。
-- 仓库不包含 dataset-local media projection。
-- OS-Atlas 属于统一 catalog，不存在单独副本。
-- `manifests/files.jsonl` 覆盖除其自身以外的全部发布对象，每个 canonical
-  仓库路径恰好对应一行。
-
-Hugging Face Xet 可以在内部进行 chunk-level dedup；这不改变上述路径级存储
-约束。
+工具会校验 mapping 引用的文件，并生成 `metadataset-full.yaml` 所需的 TAR
+索引和 Energon 元数据。
 
-## 许可与上游条款
+## License
 
-LocateAnything-Data 汇集了来自多个上游数据集的标注和引用，因此本仓库不选择、
-也不暗示一个适用于全部组成部分的统一许可证。每个上游数据集仍受其各自的
-许可证、使用条款、访问条件和署名要求约束。下载、hydrate、使用或再分发任何
-组成部分前，使用者必须查阅本 README 中链接的官方上游来源并遵守适用条款。
-本仓库不授予任何额外的上游媒体权利。
+LocateAnything-Data 由多个上游数据集组成。每个组成部分仍受其原始来源的
+license、使用条款、访问要求和署名要求约束。本仓库不授予额外的上游图片权利。
+下载、使用或再分发某个组成部分前，请查阅并遵守相应的上游条款。
 
 ## 引用