README.md

---
license: cc-by-4.0
task_categories:
- object-detection
- image-to-text
- zero-shot-image-classification
language:
- en
size_categories:
- 1M<n<10M
pretty_name: "STRAP: STructured Region Annotations single Pass"
configs:
- config_name: cc3m
  data_files:
  - split: train
    path: cc3m/train-*.parquet
- config_name: datacomp
  data_files:
  - split: train
    path: datacomp/train-*.parquet
tags:
- region-annotations
- bounding-box
- attributes
- dense-captioning
- cc3m
- datacomp
- synthetic-annotations
extra_gated_prompt: "This dataset is released for research purposes only. By requesting access, you agree to: (1) cite the associated paper in any publication or project that uses this dataset; (2) not redistribute the dataset or its annotations without permission from the authors."
extra_gated_fields:
  I agree to the terms above: checkbox
---

# STRAP: STructured Region Annotations single Pass

STRAP is a large-scale source of structured region supervision for
vision-language models. It contains annotations for **2,000,000 web images**:
1M from [Conceptual Captions 3M (CC3M)](https://ai.google.com/research/ConceptualCaptions/)
and 1M from [DataComp-1B](https://www.datacomp.ai/).
Across both subsets, STRAP describes **10.35 million objects**.

Each object connects a normalized bounding box to a basic-level label, naming
hierarchy, short description, attributes, visible parts, and categories the
object could be mistaken for. Objects cross-reference each other by ID in
interactions and occlusion relations; readable text is annotated both on
objects and at scene level. This supports region-text representation
learning, open-vocabulary recognition, attribute and part supervision,
relation modeling, and dense captioning without requiring users to run the
annotation model themselves.

All annotations for an image were generated together in one structured pass
of the multimodal large language model (MLLM)
[**Qwen3.6-35B-A3B**](https://huggingface.co/Intel/Qwen3.6-35B-A3B-int4-mixed-AutoRound).

For more details, see the dataset [**website**](https://klarajanouskova.github.io/STRAP/).

If you find the dataset useful, please cite us - an arxiv paper is coming soon :) We would also love to hear about your experience and use cases!

<p align="center">
  <img src="assets/annotation_pipeline.png" alt="STRAP pipeline: a frozen MLLM sees each image once with a fixed prompt and JSON schema, and returns one structured record per image with image-level fields and one detailed card per object" width="100%"/>
</p>

> **Images are not included.** STRAP contains annotations only, with no image
> bytes. Obtain images from the original distributions and join them using the
> provided keys; see [Getting the images](#getting-the-images).

## Dataset Summary

| | `cc3m` | `datacomp` |
|---|---|---|
| Images (rows) | 1,000,000 | 1,000,000 |
| Annotated objects | 6,229,195 | 4,121,158 |
| Objects per image (mean) | 6.2 | 4.1 |
| Image source | CC3M | DataComp-1B |
| Primary join key | `tsv_index` (row in Google's CC3M TSV) | `uid` (DataComp-1B sample ID) |
| MLLM annotator | [Qwen3.6-35B-A3B](https://huggingface.co/Intel/Qwen3.6-35B-A3B-int4-mixed-AutoRound) | [Qwen3.6-35B-A3B](https://huggingface.co/Intel/Qwen3.6-35B-A3B-int4-mixed-AutoRound) |

Created and released by the Visual Recognition Group FEE CTU in Prague.
Annotations are released under CC BY 4.0.

## Quick Start

Stream the dataset to inspect records without first downloading an entire
config:

```python
from datasets import load_dataset

cc3m = load_dataset(
    "vrg-prague/STRAP", "cc3m", split="train", streaming=True
)
example = next(iter(cc3m))
```

Use the `datacomp` config name for the DataComp subset. Omit `streaming=True`
to download and load a complete config.

## Example

The annotations of one image (bounding boxes and basic labels):

<p align="center">
  <img src="assets/example_cc3m.jpg" alt="STRAP example: a squirrel eating a nut in a snowman teacup, with all annotated bounding boxes" width="70%"/>
</p>

The same annotation as a record (abbreviated: one of its seven objects and
two of its four interactions shown, floats rounded; every object carries the
full structure):

```jsonc
{
  "tsv_index": 1730823,
  "url": "https://i.pinimg.com/736x/a1/03/88/a1038843468be4725af63cabb2532d4e--secret-squirrel-winter-scenes.jpg",
  "original_caption": "i 'm going nuts over this squirrel",
  "width": 600, "height": 451,
  "cc3m_wds_key": "001779031",
  "image_summary": "A close-up photograph of a squirrel sitting inside a large, blue, snowman-themed teacup, eating a nut against a snowy background.",
  "scene_context": "A close-up outdoor scene with a shallow depth of field, featuring a squirrel in a decorative cup against a blurred snowy background.",
  "objects": [
    {
      "id": 1,
      "basic_label": "squirrel",
      "salience": 0.95,
      "brief_description": "A squirrel sitting inside a teacup, eating a nut.",
      "bounding_box": [0.33, 0.10, 0.85, 0.72],
      "hierarchy": {
        "coarser_levels": [{"label": "mammal", "confidence": 0.99}, {"label": "rodent", "confidence": 0.99}],
        "basic_confidence": 0.99,
        "finer_levels": [{"label": "red squirrel", "confidence": 0.85}],
        "confidence_note": "The reddish fur on the tail and back strongly suggests a red squirrel, though the greyish body is also common in some populations or lighting."
      },
      "intrinsic_parts": ["head", "torso", "left_front_paw", "right_front_paw", "bushy_tail", "whiskers", "ears"],
      "attributes": {
        "colors": ["grey", "brown", "rust-red"],
        "materials": ["fur"],
        "pose_or_state": "sitting, eating",
        "other": ["long whiskers", "large dark eye"]
      }
      // ... visibility, anchor_point, count, confusable_with, context_typicality, ...
    }
    // ... 6 more objects (nut, teacup, saucer, ...)
  ],
  "interactions": [
    {
      "object_ids": [1, 2],
      "kind": "eating",
      "direction": {"from_object_id": 1, "to_object_id": 2, "is_directional": true},
      "note": "The squirrel (1) is holding and eating the nut (2)."
    },
    {
      "object_ids": [1, 3],
      "kind": "sitting_in",
      "direction": {"from_object_id": 1, "to_object_id": 3, "is_directional": true},
      "note": "The squirrel (1) is sitting inside the teacup (3)."
    }
    // ... 2 more interactions
  ]
}
```

## Dataset Structure

STRAP has two configs (`cc3m`, `datacomp`), each with one `train` split stored
as 20 parquet shards. One row represents one image.

### Row fields

| column | type | description |
|---|---|---|
| `tsv_index` | int | *(cc3m only)* stable 0-based row join key into Google's original `Train_GCC-training.tsv` |
| `uid` | string | *(datacomp only)* DataComp-1B's canonical sample ID (`md5(text + url)`), matches the `uid` column of the official metadata |
| `url` | string | original image URL |
| `original_caption` | string | original web alt-text caption: the caption from the official CC3M TSV / the `text` field of the DataComp-1B metadata, reproduced exactly |
| `width`, `height` | int | pixel size of the annotated image |
| `cc3m_wds_key` | string | *(cc3m only)* `__key__` in the hosted `pixparse/cc3m-wds` mirror (unique per row) |
| `image_summary` | string | one-sentence summary of the image |
| `scene_context` | string | scene/setting description |
| `objects` | list&lt;struct&gt; | annotated objects, see below |
| `interactions` | list&lt;struct&gt; | `object_ids`, `kind`, `direction{from_object_id, to_object_id, is_directional}`, `note` |
| `global_text` | list&lt;struct&gt; | readable text not tied to an object: `text`, `location`, `legibility` |
| `uncertain_items` | list&lt;struct&gt; | `description`, `possible_classes`, `reason` |
| `annotation_json` | string | the verbatim model output; a lossless superset of the typed columns |

### Object fields

Each element of `objects` (a `?` after a type marks a nullable field):

| field | type | description |
|---|---|---|
| `id` | int | object ID, referenced by `interactions`, `occluded_by_id`, `associated_items.object_id` |
| `basic_label` | string | basic-level category name |
| `hierarchy` | struct? | `coarser_levels`/`finer_levels` (lists of `{label, confidence}`), `basic_confidence`, `confidence_note`; null for low-salience objects |
| `salience` | float | 0 to 1; ≥ 0.5 is the schema's "salient object" threshold |
| `brief_description` | string | one-line free-text description |
| `bounding_box` | [float]×4 | normalized `[x0, y0, x1, y1]` |
| `anchor_point` | [float]×2 | normalized representative point |
| `bounding_description` | string? | textual location ("upper left corner", …) |
| `count` | int? | instance count for group boxes |
| `group_id` | int? | shared ID when several rows form one group |
| `visibility` | string | `fully_visible` / `partially_occluded` / … |
| `occluded_by_id` | int? | ID of the occluding object |
| `rendition` | struct? | `{medium, depicts}` when the object depicts something (statue, drawing, screen, …) |
| `confusable_with` | list | `{label, reason}`: categories the object could be mistaken for, each with a reason |
| `context_typicality` | struct? | `{is_typical_context, context_note}` |
| `intrinsic_parts` | list&lt;string&gt; | structural parts (head, wheel, …) |
| `associated_items` | list | `{item, relation, covers, object_id}`: worn/carried/mounted items |
| `attributes` | struct | `colors`, `materials`, `pose_or_state`, `size_cue`, `expression`, `markings`, `clothing`, `condition`, `other` |
| `text_on_object` | list | `{text, location_on_object, legibility}` |
| `recognition_requires_text` | bool | label depends on reading text |
| `recognition_requires_context` | bool | label depends on scene context |

Nullable nested structs are `null` when absent (e.g. `hierarchy` for
low-salience objects). `annotation_json` preserves the raw model record
verbatim.

### Coordinate convention

`bounding_box` = `[x0, y0, x1, y1]` and `anchor_point` = `[x, y]`, both
**normalized to [0, 1]** relative to the full image. They remain valid at any
image resolution. `width` and `height` are
the pixel dimensions of the image the annotator saw; use them to map a box
into that image's pixel coordinates
(`pixel_box = [x0·W, y0·H, x1·W, y1·H]`) and to check whether a downloaded
image matches the annotated version.

## Getting the images

### `cc3m` config

- **Official CC3M** ([Google's TSV release](https://ai.google.com/research/ConceptualCaptions/download)):
  join on `tsv_index` (0-based row index of `Train_GCC-training.tsv`) or on
  `url`.
- **Hosted images** ([`pixparse/cc3m-wds`](https://huggingface.co/datasets/pixparse/cc3m-wds)):
  join on `cc3m_wds_key` (the WebDataset `__key__`). The easiest way to
  obtain the images themselves.

### `datacomp` config

- **Official DataComp-1B** ([metadata](https://huggingface.co/datasets/mlfoundations/datacomp_1b)):
  join on `uid` (the metadata's sample ID), or download images directly from
  `url` (e.g. with [img2dataset](https://github.com/rom1504/img2dataset)).

## Annotation Process

For each image, **Qwen3.6-35B-A3B** (the
[Intel int4-mixed AutoRound quantization](https://huggingface.co/Intel/Qwen3.6-35B-A3B-int4-mixed-AutoRound))
received the image, a fixed instruction prompt, and a fixed JSON schema. Source captions were not provided to the
annotator. The model produced the image-level and object-level fields together
in one response, which keeps boxes, descriptions, attributes, and relations
connected through shared object IDs.

Records were checked for successful parsing and schema conformance. This
validation guarantees a consistent machine-readable structure, not semantic
correctness; labels, descriptions, and localization remain model-generated.

## Considerations for Using the Data

### Intended Use

STRAP is intended for training and analyzing region-level vision-language
representations, including region-text alignment, open-vocabulary recognition,
dense captioning, region attributes and parts, object relations, and structured
MLLM annotation at scale.

### Not Intended For

STRAP is not a substitute for human-verified detection ground truth. In
particular, it should not be used as the sole reference for precise localization
evaluation, calibrated confidence evaluation, or for making judgments about
individuals depicted in the images.
It also does not provide an image distribution; users must obtain images under
the terms of the source datasets.

### Limitations

Annotations are model-generated, so labels, boxes, descriptions, and
attributes may contain errors. Bounding boxes are MLLM estimates rather than
detector outputs and should be treated as approximate localization. Confidence
fields reflect the annotator's self-assessment, not calibrated probabilities.
The underlying CC3M and DataComp-1B images are web-scale data and inherit the
biases of those collections.

## Licensing

- **Generated annotations** (`image_summary`, `scene_context`, `objects`,
  `interactions`, `global_text`, `uncertain_items`, `annotation_json`):
  **CC BY 4.0**. They were generated with
  [Qwen3.6-35B-A3B](https://huggingface.co/Qwen/Qwen3.6-35B-A3B), released
  under Apache 2.0.
- **Source and join metadata** (`url`, `original_caption`, `width`,
  `height`, `tsv_index`, `cc3m_wds_key`, `uid`) is redistributed solely to
  enable joining: CC3M metadata under the
  [Conceptual Captions terms](https://github.com/google-research-datasets/conceptual-captions)
  (attribution of the source required), and DataComp-1B metadata under
  **CC BY 4.0**.
- **Images** are *not* distributed here. CC3M images/URLs are subject to the
  [Conceptual Captions terms](https://github.com/google-research-datasets/conceptual-captions);
  DataComp-1B content is subject to its source terms.

## Dataset Card Contact

[Visual Recognition Group FEE CTU in Prague](https://huggingface.co/vrg-prague)
([group website](https://vrg.fel.cvut.cz))