EgoSurgHands contains egocentric video frames of consenting surgical trainees performing training exercises on synthetic-skin pads. It is released under CC BY-NC-SA 4.0, and the Project Aria research license terms also apply (https://www.projectaria.com/datasets/dataset-license-agreement/). Your Hugging Face username and email address are shared with the authors when you submit this form.

EgoSurgHands

An egocentric 3D hand-pose dataset and benchmark for open-surgery training (NeurIPS 2026, Evaluations & Datasets Track).

EgoSurgHands contains real egocentric recordings of surgical trainees performing standard open-surgery training tasks (suturing, knot tying, needle loading, cutting) on synthetic-skin training pads, captured with Meta Project Aria glasses. Each row is one hand in one frame: the RGB frame paired with a 21-joint 3D hand label derived from Aria Machine Perception Services (MPS).

To our knowledge it is the first benchmark for real-captured egocentric hand pose estimation in open-surgery training on synthetic pads. Evaluation labels are sensor-fused, human-validated pseudo-ground truth. They come from Aria MPS, which fuses cues that no monocular RGB model can access. Every evaluation frame passed human review with examiner cross-review and senior-surgeon adjudication of flagged cases, and an external multi-view triangulated reference checks the labels. They are not mocap-grade.

Paper: EgoSurgHands: An Egocentric 3D Hand Pose Dataset & Benchmark for Open-Surgery Training, NeurIPS 2026, Evaluations & Datasets Track.
Code and leaderboard: https://egosurghands.github.io
Version: 1.0.0, the release reviewed for NeurIPS 2026 (see Changelog).
Access: gated. Submit the short request form on this page (name, affiliation, intended use, and agreement to the terms). The authors review each request.

At a glance

	Train	Validation (evaluation)
Hand-instance rows	41,909	9,476
Procedure recordings (PRs)	55	13
Unique frames	26,087	6,559
Wearers	4	4
Glove colours	3	3
Tasks	11 (+560 rows without a task label)	13 (one per PR)

Splits: train and validation are PR-disjoint, with no shared frames. They share wearers and protocol, so the original adaptation split is within-corpus. The paper adds a leave-one-wearer-out evaluation that tests unseen trainees.
Capture: Beth Israel Deaconess Medical Center, Boston. Four trainees (P1-P4), from first-year medical student to third-year resident, in three glove colours.
Frames: 1408×1408 RGB from the Aria RGB camera (fisheye lens), JPEG, sampled at 5 Hz (every 6th frame of the 30 fps stream).

Quick start

Request access on this page, then log in once with huggingface-cli login (or pass token=...).

import io
import numpy as np
from datasets import load_dataset
from PIL import Image

ds = load_dataset("harvardsil/EgoSurgHands", split="validation")
row = ds[0]
img = Image.open(io.BytesIO(row["image"]))           # JPEG bytes, 1408x1408, stored orientation (see below)
j3d = np.asarray(row["joints_3d"]).reshape(21, 3)    # benchmark label, metres, RGB camera frame
upright = np.rot90(np.asarray(img), 2)               # natural viewing orientation

Conventions you need

Joint order. The arrays use OpenPose-21 order (wrist 0, thumb 1-4, index 5-8, middle 9-12, ring 13-16, little 17-20), with one exception. Slot 1 holds the MPS palm-centre landmark, not a thumb CMC joint, because MPS has no thumb CMC. The benchmark metric uses the other 20 joints (OP_VALID = [0, 2, 3, ..., 20]).
Flat arrays. joints_3d (63 values), joints_2d (42), intrinsics (9) and extrinsics (16) are stored flat in row-major order. Reshape them to 21×3, 21×2, 3×3 and 4×4.
Image orientation. Frames are stored rotated 180° from natural viewing (the wearer's arms enter from the top). Use np.rot90(img, 2) for display.
joints_2d is not image pixels. It is a pinhole projection of joints_3d with f = 606.5 and c = 704 (the intrinsics matrix), with no lens distortion and no rotation. It is the "virtual pinhole" space of the paper's PA-P2D metric, and bbox is in the same space.
Timestamps. frame_ts_us is on the Aria device clock (µs since device start), not Unix time.

Columns

Column	Type	Description
`frame_key`	string	`PR{nnn}/{frame_id:05d}_{hand_side}`
`source_tag`	string	`slam` if MPS SLAM succeeded for the recording, otherwise `noslam` (training only; every evaluation recording is `slam`)
`image`	binary	JPEG bytes of the 1408×1408 frame, stored orientation
`image_width`, `image_height`	int32	1408
`intrinsics`	float32[9]	3×3 pinhole matrix used by `joints_2d` (f = 606.5, c = 704); identical in every row
`extrinsics`	float32[16]	4×4 identity
`bbox`	float32[4]	`[x0, y0, x1, y1]`: hull of `joints_2d` plus 16.5% padding, clipped to the frame (pinhole space)
`hand_side`	string	`left` or `right`
`joints_3d`	float32[63]	Benchmark label. 21×3 joints in metres, RGB camera frame, from Aria MPS
`joints_2d`	float32[42]	21×2 pinhole projection of `joints_3d` (see convention 4)
`pr`, `sequence_name`	string	procedure recording id, e.g. `PR108`
`frame_id`	int32	frame index in the 5 Hz sequence
`frame_ts_us`	int64	image capture time, Aria device clock (µs)
`wearer_id`	string	de-identified wearer `P1`-`P4`
`experience`	string	training stage: `MS1`, `PGY0` or `PGY3`
`handedness`	string	wearer's dominant hand
`glove_color`	string	`blue`, `white` or `brown`
`task`	string	surgical task; 560 training rows from three recordings carry `incomplete`, `no task` or `none`
`recording_day`	string	`day_1` or `day_2` (real dates removed)

How the labels were made

Capture. Aria glasses, 30 fps RGB. Each recording was processed with Meta's cloud-based MPS, whose hand tracking gives 21 landmarks per hand at 30 Hz with a per-hand confidence.
Frames. Frames are sampled at 5 Hz.
Sampling the label. MPS landmarks were linearly interpolated and kept only if both bracketing samples had per-hand confidence ≥ 0.97. Rows without a valid label were dropped (1,047 training rows, no validation rows).
Transform to the camera frame. Landmarks were moved from the device frame into the RGB camera frame using the MPS online calibration of the same recording session, and reordered to the convention above.
Human review. Every evaluation frame went through human review: expert primary review, examiner cross-review, and senior-surgeon adjudication of flagged cases.
- Across 9,476 review decisions, 43.1% of frames were accepted as-is and 51.9% after a drag correction (median 8.2 mm; mean 9.8 mm over the 4,914 corrections; 5.3 mm over all reviewed hand instances); only 5.1% were rejected. An offset is recorded only after an annotator flags an overlay and is stored as a global image-plane translation, so these values quantify the magnitude of human intervention, not per-joint 3D error.
- On a 500-frame stratified audit subset in which three annotators independently made keep/reject decisions, pooled pairwise agreement is 72.9% (Cohen's κ 0.423, Fleiss κ 0.421, Krippendorff α 0.422).
External triangulated check. In a separate validation capture, two GoPro cameras and an overhead iPhone recorded alongside the Aria glasses while all four wearers performed the training tasks in all three glove colours. Against this multi-view triangulated reference (500 frames, 562 hands), the median per-joint deviation of the MPS labels is 15.9 mm after one similarity registration per recording and 9.8 mm after per-hand Procrustes alignment.
Reproducibility. All 51,385 labels were recomputed from the raw MPS outputs and match the stored values exactly (largest difference 0.0001 mm).

Residual benchmark error combines model error, MPS uncertainty, calibration effects and MPS-to-MANO skeletal differences. The paper characterises a label resolution scale on the order of 1 cm. Zero-shot differences across methods (15-29 mm) exceed it. Differences among corrected models below about 1 cm are not resolvable with these labels.

Benchmark protocol

Metric: PA-MPJPE on the 20 OP_VALID joints (Umeyama similarity alignment) against joints_3d on the validation split. The paper also reports MPJPE-RR, Root error, MPJPE-abs and PA-P2D.
Adapters: any adapter (for example the paper's JRN baseline) is trained on the train split only. Confidence intervals bootstrap over the 13 evaluation PRs.
Additional protocol: leave-one-wearer-out evaluation, described in the paper.
Benchmark code and leaderboard: https://egosurghands.github.io

Composition

Wearer	Training stage	Dominant hand	Gloves	Rows (train / val)
P1	PGY3	right	blue	10,780 / 4,694
P2	MS1	left	white	13,002 / 3,198
P3	PGY0	right	white (train also blue)	3,698 / 923
P4	PGY3	right	brown (train also blue)	14,429 / 661

All brown-glove evaluation rows (661) come from wearer P4 in one recording, so glove colour and wearer are confounded in that stratum.

Tasks (rows train / validation):

Task	Train	Val
simple interrupted + instrument tie	13,125	176
running sub-cuticular + aberdeen knot	3,667	2,162
horizontal mattress + one hand tie	4,852	511
needle loading	4,750	478
horizontal mattress + instrument tie	3,548	1,389
figure of 8 + instrument tie	3,141	1,331
figure of 8 + two handed tie	2,829	661
cutting	3,258	75
simple interrupted + two hand tie	818	638
figure of 8 + one hand tie	778	649
vertical mattress + instrument tie	0	952
horizontal mattress	583	161
two hand tie	0	293
no task label (`incomplete`, `no task`, `none`)	560	0

Uses

Validated:

zero-shot benchmarking of monocular hand-pose models on egocentric surgical-training video;
within-dataset MPS-label adaptation baselines such as JRN;
in-distribution per-task, per-wearer and per-glove analysis, keeping the P4/brown confound in mind.

Not validated, and not to be used for:

performance scoring for hiring or licensure decisions;
live-OR deployment without further domain-specific work;
person re-identification;
cross-site or cross-population generalisation claims;
fine-grained ranking of corrected models whose differences fall below the roughly 1 cm label resolution scale.

Limitations

Scope: four wearers at one institution, on synthetic-skin training pads rather than in live surgery. Egocentric monocular RGB.
Labels: sensor-fused, human-validated pseudo-ground truth, not mocap-grade.

Ethics and privacy

Participants: consenting adult trainees on synthetic-skin pads. No patients and no protected health information. Audio is not released.
De-identification: wearers are de-identified as P1-P4, recording dates are replaced by day_1/day_2, and timestamps use the device clock.
Images: no faces, name tags or clinical-systems content are captured.
IRB: because the dataset contains no patient data and only consenting adult participants performing routine surgical-training exercises on synthetic-skin pads, formal IRB review was not required under the institution's non-human-subjects-research policy.
Access: gated per person. Requesters agree to non-commercial use, to the Project Aria terms, not to identify the participants, and not to score individual trainees.

License

CC BY-NC-SA 4.0 (non-commercial, share-alike, attribution). The Project Aria research license terms also apply: https://www.projectaria.com/datasets/dataset-license-agreement/

Citation

@inproceedings{pal2026egosurghands,
  title     = {{EgoSurgHands}: An Egocentric {3D} Hand Pose Dataset \& Benchmark for Open-Surgery Training},
  author    = {Pal, Ankit and Kugener, Guillaume and Volpi, Sofia and Schwartz, Joshua and Idarraga, Alex and Rajpurkar, Pranav and Brat, Gabriel A.},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Evaluations and Datasets Track},
  year      = {2026}
}

Authors: Ankit Pal (Department of Biomedical Informatics, Harvard Medical School); Guillaume Kugener, Sofia Volpi, Joshua Schwartz and Alex Idarraga (Department of Surgery, Beth Israel Deaconess Medical Center); Pranav Rajpurkar (Department of Biomedical Informatics, Harvard Medical School) and Gabriel A. Brat (Department of Biomedical Informatics, Harvard Medical School; Department of Surgery, Beth Israel Deaconess Medical Center), equal senior authorship.

Acknowledgments

We thank the surgical trainees who volunteered to wear the Aria glasses during their training sessions, and the annotators and surgeons who reviewed, corrected and adjudicated the evaluation frames. We thank the Department of Surgery at Beth Israel Deaconess Medical Center for hosting the recordings and providing the training materials, and the Project Aria team at Meta Reality Labs Research for the Aria Research Kit and Machine Perception Services, which produce the hand landmarks behind every label. We thank Nebius for the GPU compute that powered our zero-shot evaluation of all eight methods, the training of the JRN baselines across every split and wearer fold, and the leave-one-wearer-out study.

Maintenance

Maintained by the authors. Hosted on Hugging Face for at least 24 months from publication; a DOI will be added to this card. Versions follow semantic versioning, and 1.x keeps the benchmark rows and labels fixed.

Files

train/ (12 shards), validation/ (3 shards): Parquet.
DATASHEET.md: Datasheet for Datasets.
croissant.json: Croissant 1.1 metadata with responsible-AI fields.
BUILD_MANIFEST.json: row counts and provenance.

Changelog

1.0.0 (2026-05-07): the release reviewed for NeurIPS 2026. The documentation was updated on 2026-11-06 for the camera-ready paper (authors, links, access form and column descriptions); the data files are unchanged.

Downloads last month: 93

Total file size:

10.8 GB