HandNet

214,971 annotated hand depth frames, with six measured sensor poses per frame. HandNet supports research on fingertip localization, hand segmentation and orientation normalization. The original data was recorded at the Technion GIP lab using an Intel RealSense camera and TrakStar magnetic sensors.

Explore real samples in 2D and 3D · Paper · Original MATLAB repository · Data dictionary

Open release

Aaron Wetzler releases the dataset in this repository under Creative Commons Attribution 4.0 International (CC BY 4.0) on November 10, 2026. Reuse, adaptation and commercial use are permitted under that license, including attribution and marking changes. Attribute HandNet to Aaron Wetzler, Ron Slossberg and Ron Kimmel, link this release and the license, and cite the paper below. The full legal terms are in LICENSE.

This is a new open license for this dataset release. The linked historical MATLAB repository contains an academic-use-only notice. Its source code is not mirrored or relicensed here. Newly written inspection utilities in tools/ have their own MIT license.

Files and splits

All seven distribution archives are preserved with their original filenames and bytes. They contain MATLAB .mat examples; the additional transformed archives reuse examples rather than adding new independent recordings.

File under archives/	Bytes	GiB	Contents
TrainData.rar	12,977,627,177	12.086	202,198 original training examples
TestData.zip	642,222,352	0.598	10,000 original test examples
ValidationData.zip	178,069,511	0.166	2,773 original validation examples
TrainData_DerotGT.zip	10,846,508,573	10.102	Training examples normalized with ground-truth orientation
ValidationData_DerotGT.zip	148,671,962	0.138	Validation examples normalized with ground-truth orientation
TestData_Perturbed.zip	534,164,788	0.497	Test examples with random in-plane rotation
ValidationData_Perturbed.zip	148,312,479	0.138	Validation examples with random in-plane rotation

Seven-archive download total: 25,475,576,842 bytes (23.726 GiB). Only the three original splits count toward the 214,971 unique released examples. The original split is a shuffled frame split, not a claimed held-out-subject benchmark. Preserve the supplied split membership when comparing results.

The 2015 paper describes an earlier total of 212,928 frames and a 202,928/10,000 train/test split. The released archive counts and original viewer instead specify 202,198 training, 10,000 test and 2,773 validation frames. This card describes the released files; those historical counts should not be silently substituted for one another.

DerotGT is an oracle transformation using measured palm orientation. It is not a model prediction. The perturbed variants apply uniformly sampled rotations between -90 and +90 degrees, consistently transforming images and camera-space annotations. See the original archive descriptions.

Small browser preview

The Hub viewer uses preview/samples.parquet: 96 derived visualization rows (3,306,503 bytes), selected from original and transformed examples. This preview is not the complete dataset or a benchmark split. It embeds contrast-rendered depth and infrared images, colorized labels, six sensor poses and source frame identifiers. These display images do not preserve raw depth or label-array values. The full 214,971 original frames remain in the archives above.

python -m pip install datasets pillow

from datasets import load_dataset

preview = load_dataset("twerdster/HandNet", "preview", split="preview")
print(len(preview))  # 96 visualization rows, including transformed counterparts
row = preview[0]
print(row["source_archive"], row["source_member"])
print(row["sensor_names"], row["sensor_xyz_mm"])

Preview image columns are embedded PNGs using the Hugging Face Image feature with decoding disabled. Read their bytes with an image reader, or cast the columns with datasets.Image(decode=True). See samples.schema.json for the complete preview schema and tools/README.md to reproduce it.

Download and read an original frame

Download only the files you need. These commands initiate downloads when you run them:

python -m pip install huggingface_hub numpy scipy
hf download twerdster/HandNet archives/ValidationData.zip --repo-type dataset --local-dir ./handnet

Or use a direct file URL:

wget -c https://huggingface.co/datasets/twerdster/HandNet/resolve/main/archives/ValidationData.zip

You can read a ZIP member without extracting the full archive:

from pathlib import PurePosixPath
from zipfile import ZipFile
import numpy as np
from scipy.io import loadmat
from huggingface_hub import hf_hub_download

archive = hf_hub_download(
    repo_id="twerdster/HandNet",
    filename="archives/ValidationData.zip",
    repo_type="dataset",
)
with ZipFile(archive) as z:
    members = sorted(
        n for n in z.namelist()
        if PurePosixPath(n).name.startswith("Data")
        and n.lower().endswith(".mat")
        and "__MACOSX" not in PurePosixPath(n).parts
    )
    with z.open(members[0]) as f:
        frame = loadmat(f)

depth_mm = frame["depth"]       # 240 rows x 320 columns
sensor_xyz_mm = frame["pos"]  # 3 x 6: thumb through little finger, then palm
sensor_rotations = frame["rot"]  # 3 x 3 x 6
print(depth_mm.shape, sensor_xyz_mm.shape, sensor_rotations.shape)

The original training file is RAR, not ZIP. Use a RAR-capable extractor to read TrainData.rar, then load an extracted .mat with scipy.io.loadmat. The supplied inspection utility accepts extracted .mat files and ZIP members; it does not require the historical MATLAB source.

The complete MATLAB archives are not converted into a datasets.load_dataset() table; the preview configuration covers only the 96 display rows. Use the archive reader above for native arrays. calibration.json provides the original camera matrix for point-cloud reconstruction.

Provenance and integrity

The source of this release is the existing local copy of OpenDataLab/HandNet, whose archive family matches the original authors' distribution list. The seven archive bytes are rehosted unchanged, without recompression or annotation edits. Their origin is not asserted to be the HyperAI version-1 mirror.

The release includes SHA256SUMS and a machine-readable manifest.json with archive byte sizes and SHA-256 hashes. After downloading a complete local release, verify it from the repository root:

sha256sum -c SHA256SUMS
# macOS alternative:
shasum -a 256 -c SHA256SUMS

Hash agreement verifies equality with this release; it does not establish equality with a currently unavailable original server copy. The dataset does not include the separately distributed original result archives or trained model weights.

Scope and limitations

The data represents ten participants recorded in a laboratory. Frame diversity does not imply broad population, environment, device or gesture coverage. Temporal neighbours can occur in different shuffled splits; a high score on these splits does not establish unseen-subject generalization.

Annotations measure five fingertips and a palm sensor, including orientation. These are not a full anatomical hand-joint skeleton. Magnetic measurements and generated label/heatmap images have calibration, placement, projection and visibility limitations. Out-of-frame sensors can produce non-finite heatmap values; missing depth is zero. Inspect fields and masks before training.

The inspected image modalities are depth and monochrome infrared (ir); this release does not promise RGB images. The rgbd discovery tag reflects the capture-device family. The Explorer's displayed annotations and ground-truth derotation are data visualizations, not newly evaluated detector predictions.

Citation

@inproceedings{WetzlerBMVC15,
  author = {Aaron Wetzler and Ron Slossberg and Ron Kimmel},
  title = {Rule Of Thumb: Deep Derotation for Improved Fingertip Detection},
  booktitle = {Proceedings of the British Machine Vision Conference (BMVC)},
  year = {2015},
  pages = {33.1--33.12},
  doi = {10.5244/C.29.33},
  url = {https://arxiv.org/abs/1507.05726}
}