ryushinn/11k-Hands-BBox-Keypoint

---
license: mit
task_categories:
- image-to-text
tags:
- hand-pose-estimation
- keypoints
- bounding-boxes
- rtmpose
- mmpose
- 11k-hands
pretty_name: 11k-Hands BBox Keypoint
size_categories:
- 10K<n<100K
---

# 11k-Hands BBox Keypoint

This dataset contains estimated hand bounding boxes and 21-point hand keypoints for [`ryushinn/11k-Hands`](https://huggingface.co/datasets/ryushinn/11k-Hands). It is an annotations-only sidecar dataset: it does **not** duplicate the source images.

Rows preserve the same split and row order as the source dataset. To pair an annotation row with its image, load the same split from `ryushinn/11k-Hands` and use the same row index.

## Dataset structure

The dataset has one split:

| Split | Rows |
| --- | ---: |
| `train` | 11,076 |

Each row has exactly four columns:

| Column | Type / shape | Description |
| --- | --- | --- |
| `bboxes_xyxy` | `float32[4]` | Highest-confidence detected hand box as `[x1, y1, x2, y2]` in source-image pixel coordinates. |
| `bbox_scores` | `float32` | Confidence score for the selected hand bounding box. |
| `keypoints_xy` | `float32[21][2]` | Estimated 21-point hand keypoints in source-image pixel coordinates. |
| `keypoint_scores` | `float32[21]` | Confidence score for each keypoint. |

## Keypoint order

The 21 keypoints follow the COCO-WholeBody hand convention used by MMPose:

| Index | Name |
| ---: | --- |
| 0 | wrist |
| 1 | thumb1 |
| 2 | thumb2 |
| 3 | thumb3 |
| 4 | thumb4 |
| 5 | forefinger1 |
| 6 | forefinger2 |
| 7 | forefinger3 |
| 8 | forefinger4 |
| 9 | middle_finger1 |
| 10 | middle_finger2 |
| 11 | middle_finger3 |
| 12 | middle_finger4 |
| 13 | ring_finger1 |
| 14 | ring_finger2 |
| 15 | ring_finger3 |
| 16 | ring_finger4 |
| 17 | pinky_finger1 |
| 18 | pinky_finger2 |
| 19 | pinky_finger3 |
| 20 | pinky_finger4 |

Skeleton edges used for visualization:

```python
[
(0, 1), (1, 2), (2, 3), (3, 4),
(0, 5), (5, 6), (6, 7), (7, 8),
(0, 9), (9, 10), (10, 11), (11, 12),
(0, 13), (13, 14), (14, 15), (15, 16),
(0, 17), (17, 18), (18, 19), (19, 20),
]
```

## Model configuration

Annotations were generated with OpenMMLab models:

- Hand detector: **RTMDet-Nano hand detector**
- Config: `rtmdet_nano_320-8xb32_hand.py`
- Checkpoint: `rtmdet_nano_8xb32-300e_hand-267f9c8f.pth`
- Hand pose estimator: **RTMPose-M Hand5**
- Config: `rtmpose-m_8xb256-210e_hand5-256x256.py`
- Checkpoint: `rtmpose-m_simcc-hand5_pt-aic-coco_210e-256x256-74fb594_20230320.pth`

The RTMPose-M Hand5 model is trained on a mixture of hand datasets including COCO-WholeBody-Hand, OneHand10K, FreiHand2D, RHD2D, and Halpe hand annotations.

Reference: Jiang et al., **RTMPose: Real-Time Multi-Person Pose Estimation based on MMPose**, arXiv:2303.07399.

## Usage

```python
from datasets import load_dataset, concatenate_datasets

images = load_dataset("ryushinn/11k-Hands", split="train")
ann = load_dataset("ryushinn/11k-Hands-BBox-Keypoint", split="train")

ds = concatenate_datasets([images, ann], axis=1)
idx = 0
image = ds[idx]["image"]
bbox = ds[idx]["bboxes_xyxy"]
keypoints_xy = ds[idx]["keypoints_xy"]
keypoint_scores = ds[idx]["keypoint_scores"]
```

## Visualization previews

The following preview images show source images with the estimated hand bounding box and keypoint skeleton overlayed.

| Row 0 | Row 1 |
| --- | --- |
|  |  |

| Row 5538 | Row 11075 |
| --- | --- |
|  |  |

## Notes

These annotations are model-estimated pseudo-labels, not manual ground-truth annotations. They are intended for research workflows where reproducible hand bounding boxes and hand pose estimates are useful alongside the original `ryushinn/11k-Hands` images.