Datasets:

BreakpointAI
/

breakpoint-grounding-5m

---
dataset_info:
  features:
  - name: object_captions
    list: string
  - name: image_caption
    dtype: string
  - name: normalized_boxes
    list:
      list: float32
  - name: img_size_wh
    list: int32
  - name: source_dataset
    dtype: string
  - name: relative_path
    dtype: string
  splits:
  - name: train
    num_bytes: 10259224594
    num_examples: 5788300
  download_size: 8572776179
  dataset_size: 10259224594
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
language:
- en
license: cc-by-4.0
size_categories:
- 1M<n<10M
task_categories:
- object-detection
pretty_name: Breakpoint Grounding 5M
tags:
- object-detection
- bounding-boxes
- grounding
- image-captioning
---

# Breakpoint Grounding 5M

## Quick start

```python
from datasets import load_dataset
ds = load_dataset("BreakpointAI/breakpoint-grounding-5m", split="train")
ds[0]  # {'source_dataset': ..., 'relative_path': ..., 'image_caption': ..., 'object_captions': [...], 'normalized_boxes': [...], 'img_size_wh': [...]}
```

## Dataset summary

An annotations-only companion to
[Breakpoint Grounding 55M](https://huggingface.co/datasets/BreakpointAI/breakpoint-grounding-55m):
image-level captions, object bounding boxes, and a caption for every box, for **5,788,300 images**
drawn from 16 public object-detection/tracking benchmarks.

In contrast to the 55M dataset, where the bounding box annotations are model-generated, in this dataset the bounding box annotations here are hand-labeled. 

**No image bytes are included in this release.** Several of the source benchmarks (notably
ImageNet and Objects365) restrict redistribution of their images. To stay clear of that, this
dataset publishes only the annotations Breakpoint authored, plus a `source_dataset` +
`relative_path` pointer back to each image — usable by anyone who already has their own licensed
copy of the relevant source dataset, laid out in its standard directory structure.

- **5,788,300 images**, each with one image-level caption
- **51,196,466 bounding boxes** (8.84 per image on average), each with its own caption

## Source datasets

| Source | Share | Description |
|--------|------:|-------------|
| Open Images (train/test/val) | 34.7% | Google's large-scale detection/annotation dataset |
| Objects365 (train/test/unlabeled) | 34.8% | Megvii/BAAI large-vocabulary object detection dataset |
| LaSOT | 13.3% | Large-scale single-object tracking benchmark |
| ImageNet | 10.3% | Large-scale image classification dataset (ILSVRC) |
| COCO (train) | 2.0% | Common Objects in Context detection/captioning benchmark |
| Open Images (test) | 1.9% | (see above) |
| RoboFlow 100 | 1.7% | Benchmark aggregating 100 diverse object-detection datasets |
| LVIS (train) | 1.7% | Large-vocabulary instance segmentation, built on COCO images |
| ObjectNet3D | 1.6% | Object recognition/pose dataset |
| BDD100K | 1.4% | Berkeley DeepDrive large-scale driving-scene dataset |
| CrowdHuman | 0.3% | Pedestrian / crowd detection benchmark |
| UWISOccludedDataset | 0.15% | Occlusion-focused detection dataset |
| LOCO | 0.1% | Logistics-scene object detection dataset |

<!-- TODO: verify descriptions/full names above, especially UWISOccludedDataset and LOCO, whose
     provenance we're least certain about. Percentages from an earlier per-source row count. -->

## How it was built

- **Object detection**: bounding boxes and class labels are given by the datasets, provided by human labelers.
- **Per-object captions**: [InternVL3-2B](https://huggingface.co/OpenGVLab/InternVL3-2B), served
  via LMDeploy. Each detected box was cropped (no padding) and captioned with:
  > `<image>` Describe this `<category name>` in 20 words or less and focus on its color,
  > appearance, size, and style, etc.

  where `<category name>` is the detector's class label for that box.
- **Image-level captions**: [InternVL3-8B](https://huggingface.co/OpenGVLab/InternVL3-8B),
  prompted with:
  > `<image>` Write a single, natural-language description of this image. Do not use bullet
  > points or lists. Keep it under 100 words.

Captions are **not human-generated** and contain the usual model-generated hallucinations.

## Dataset structure

| Field | Type | Description |
|-------|------|-------------|
| `source_dataset` | `string` | Which of the 16 source benchmarks this image came from. |
| `relative_path` | `string` | Path to the image within that source dataset's standard layout. |
| `image_caption` | `string` | A single natural-language caption describing the whole image. |
| `object_captions` | `Sequence(string)` | One caption per detected object. Index-aligned with `normalized_boxes`. |
| `normalized_boxes` | `Sequence(Sequence(float32))` | One bounding box per detected object: `[x_min, y_min, x_max, y_max]` (xyxy), normalized to `[0, 1]`. |
| `img_size_wh` | `Sequence(int32)` | Original image size as `[width, height]` in pixels. |

## Licensing

**The annotations in this dataset** (`image_caption`, `object_captions`, `normalized_boxes`,
`img_size_wh`) are released by Breakpoint AI under **CC BY 4.0**.

**Images are not included.** To use this dataset with actual pixels, obtain each source dataset
separately under its own license/terms and join on `source_dataset` + `relative_path`. Several
sources (ImageNet, Objects365 in particular) restrict commercial use and/or redistribution —
check the terms of whichever source(s) you need before using their images.

**Removal requests.** If you believe your content is represented in these annotations and want
the corresponding row removed, open a discussion on this repository.

## Citation

```bibtex
@misc{breakpoint_grounding_5m,
  title  = {Breakpoint Grounding 5M},
  author = {Wang, Franklin and Zhong, Desmond and Murdoch, Jamie},
  year   = {2026},
  url    = {https://huggingface.co/datasets/BreakpointAI/breakpoint-grounding-5m}
}
```