README.md

metadata

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. 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:

[
    (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

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.