NeurIPS 2026 — Evaluations & Datasets Track Alexandru I. Marinescu · Diana L. Borza · Adrian S. Dărăbant — Babeș-Bolyai University, Cluj-Napoca, Romania

HumanStereo is the first large-scale synthetic stereo dataset dedicated to human faces. It contains 417,280 calibrated stereo sessions of 128 photorealistic RenderPeople avatars, rendered in Unreal Engine 5 across 109 stereo configurations and 4 indoor environments, with exact metric depth, surface normals, subject masks and full stereo calibration for every session. The paper introduces two evaluation metrics for fine-grained facial geometry — the absolute Inlier Ratio IR(ε) and the Landmark 3D Distance Error (L3DE) — and benchmarks classical, learning-based and foundation stereo models.

📄 Paper: HumanStereo: A Benchmark for Metric Facial Depth Estimation and Its Evaluation, NeurIPS 2026 (link to proceedings/OpenReview to be added)
💻 Code (loader, metrics, benchmark): https://github.com/asergiu-ubb/HumanStereoNeurIPS2026Public
🔎 Small preview (< 4 GB): https://huggingface.co/datasets/asergiu/HumanStereoPreview
🥐 Croissant metadata: metadata.json (core + Responsible AI fields)
📜 License: CC BY-NC 4.0

What is in a session

Each session is one stereo capture of one avatar in one environment with one stereo configuration. Files are named {file_type}_{apartment_location}_{subject}_{setup_index}_{capture_index}.{ext} and stored under the subject folder.

Resource	Format	Content
`image_left`, `image_right`	512×512 PNG	Unrectified RGB renders of the left/right camera
`depth_map_phys_left/right`	NPZ, float16, 512×512	Metric depth in centimeters (Unreal Engine default)
`depth_map_norm_left/right`	NPZ, float16, 512×512	Depth normalised to [0, 1]
`normal_map_left/right`	NPZ, float16, 512×512×3	World-space surface normals
`mask_left/right`	NPZ, bool, 512×512	Avatar (1) vs. background (0)
`calibration`	OpenCV YAML	`M1 M2 D1 D2 R T R1 R2 P1 P2 E F Q ROI_L ROI_R SZ` — intrinsics, extrinsics and rectification of the left/right pair
`metadata`	JSON	`slider_camera_distance` (baseline, cm), `slider_camera_rotation_yaw` (per-camera vergence, deg), `camera_fov` (deg), `slider_avatar_distance` (cm)

Facial landmarks (478-point MediaPipe Face Mesh topology) used by the L3DE metric are detected from the images with tools/detect_landmarks.py in the code repository; they are not stored in the shards.

Stereo configurations

Baseline B ∈ {5, 6, …, 15} cm · horizontal FOV ∈ {10°, 15°, 20°, 25°, 30°} · per-camera vergence θ ∈ {0°, 1°, 2°} (the angle between the optical axes is 2θ). Of the 165 combinations, the 109 in which both eye sockets stay visible are kept. Per subject and environment, 10 capture sessions are rendered for θ = 0° and 5 for θ > 0°, each with a random head pose (yaw/pitch/roll in [−10°, 10°]) and subject distance (20–200 cm): (54×10 + 55×5) × 4 environments × 128 subjects = 417,280 sessions.

Splits (identity-disjoint)

Split	Subjects	Sessions
`train`	103	335,780
`validation`	25 (same as test)	70,600
`test`	25 (same as validation)	10,900

The 25 evaluation subjects are stratified over five age groups (5 each), 13 male / 12 female, and cover five of the six ethnicity groups in the dataset (the six Indian avatars are all in train). Split membership of every session is given in manifest/sessions.csv.

Demographics of the 128 avatars

Ethnicity: White 73 · Black 23 · Asian 18 · Indian 6 · Middle Eastern 5 · Hispanic/Latino 3. Age: <13: 12 · 13–18: 10 · 19–30: 73 · 31–50: 22 · 50+: 11. Gender: 73 female · 55 male.

Repository layout

metadata.json              Croissant 1.1 metadata (core + RAI)
manifest/sessions.csv      one row per session: split, subject, environment, setup, capture index, resource paths
shards/train/train-000000.tar …        WebDataset-style TAR shards (≈4 GB each)
shards/validation/validation-000000.tar …
shards/test/test-000000.tar …

Total size ≈ 1.2 TB. Start with the test shards (10,900 sessions) or the preview repository if you only want to evaluate.

Loading

from humanstereo import Dataset          # pip install -e . in the code repository

# metadata + manifest only (no data download)
ds = Dataset.from_huggingface("asergiu/HumanStereo", split="test")

# after downloading/extracting shards into <root>:
ds = Dataset.from_croissant("<root>", split="test")
session = ds[0]
left  = session.load_image("left")                  # (512, 512, 3) BGR uint8
depth = session.load_depth_phys("left")             # (512, 512) metric depth, cm
mask  = session.load_mask("left")                   # (512, 512) bool
calib = session.load_calibration()                  # dict of OpenCV matrices (M1, M2, R, T, R1, R2, P1, P2, Q, ...)
meta  = session.load_metadata()                     # baseline, vergence, FOV, subject distance

Run the benchmark: python benchmark/evaluate.py <root> --manifest-split test --models RAFT-Stereo --regions humanoid face.

Evaluation protocol (summary)

Metrics are computed over valid pixels of a region mask (full scene, humanoid, face, eye zones, mouth): δ-accuracy, RMSE, MAE, IR(ε) for ε ∈ {1, 2, 5, 10, 20} mm, and L3DE over eight anatomical landmark pairs (inter-ocular, eye widths, nose–chin, mouth width, bizygomatic width, nose–inner-eye corners). See the paper and benchmark/metrics.py.

Intended uses and limitations

Intended: benchmarking and training stereo depth estimation on faces; metric facial geometry for eyewear fitting and AR/VR calibration research; surface-normal prediction, face segmentation, landmark detection, calibration research and synthetic-to-real domain adaptation.

Not intended / prohibited: facial recognition, surveillance, or any application that could harm individuals; use as a substitute for clinical validation on real patient data.

Limitations: fully synthetic (domain gap: no real skin reflectance, hair, eyeglasses, sensor noise or calibration error); static avatars with neutral expression; lighting varies only through the four environments; 128 identities with the demographic skew listed above.

Ethics and licensing

The dataset contains no real personal data: all images are renders of commercially licensed RenderPeople scans; only renders and annotations are redistributed, never the 3D scans. The real optometric captures used in the paper's downstream IPD experiment are not part of this release. The dataset is released under CC BY-NC 4.0; the code is under the license stated in the code repository.

Maintenance

Maintained by Prof. Adrian-Sergiu Dărăbant, Department of Computer Science, Babeș-Bolyai University (sergiu.darabant@ubbcluj.ro). Errata will be listed here; released versions stay available with version identifiers. Planned extensions: dynamic facial expressions and controlled lighting.

Citation

@inproceedings{marinescu2026humanstereo,
  title     = {HumanStereo: A Benchmark for Metric Facial Depth Estimation and Its Evaluation},
  author    = {Marinescu, Alexandru I. and Borza, Diana L. and D{\u{a}}r{\u{a}}bant, Adrian S.},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Evaluations \& Datasets Track},
  year      = {2026}
}

Acknowledgements

Supported by the project Romanian Hub for Artificial Intelligence – HRIA, 2021–2027, MySMIS no. 351416, and by project PN-IV-P7-7.1-PTE-2024-0105, contract no. 43PTE (05/05/2025). 3D human scans courtesy of RenderPeople under their commercial license.

Downloads last month: 174

Total file size:

1.2 TB