| { |
| "@context": { |
| "@language": "en", |
| "@vocab": "https://schema.org/", |
| "sc": "https://schema.org/", |
| "cr": "http://mlcommons.org/croissant/", |
| "dct": "http://purl.org/dc/terms/", |
| "citeAs": "cr:citeAs", |
| "column": "cr:column", |
| "conformsTo": "dct:conformsTo", |
| "containedIn": "cr:containedIn", |
| "data": { |
| "@id": "cr:data", |
| "@type": "@json" |
| }, |
| "dataType": { |
| "@id": "cr:dataType", |
| "@type": "@vocab" |
| }, |
| "examples": { |
| "@id": "cr:examples", |
| "@type": "@json" |
| }, |
| "extract": "cr:extract", |
| "field": "cr:field", |
| "fileObject": "cr:fileObject", |
| "fileSet": "cr:fileSet", |
| "fileProperty": "cr:fileProperty", |
| "format": "cr:format", |
| "includes": "cr:includes", |
| "isLiveDataset": "cr:isLiveDataset", |
| "jsonPath": "cr:jsonPath", |
| "key": "cr:key", |
| "md5": "cr:md5", |
| "parentField": "cr:parentField", |
| "path": "cr:path", |
| "recordSet": "cr:recordSet", |
| "references": "cr:references", |
| "regex": "cr:regex", |
| "repeated": "cr:repeated", |
| "replace": "cr:replace", |
| "separator": "cr:separator", |
| "source": "cr:source", |
| "subField": "cr:subField", |
| "transform": "cr:transform", |
| "rai": "http://mlcommons.org/croissant/RAI/", |
| "samplingRate": "cr:samplingRate", |
| "equivalentProperty": "cr:equivalentProperty" |
| }, |
| "@type": "sc:Dataset", |
| "@id": "humanstereo", |
| "name": "HumanStereo", |
| "alternateName": [ |
| "HumanStereo: A Benchmark for Metric Facial Depth Estimation and Its Evaluation", |
| "UE5Sim RenderPeople Stereo Depth Disparity Dataset" |
| ], |
| "description": "HumanStereo is a large-scale synthetic stereo dataset dedicated to human faces, introduced at NeurIPS 2026 (Evaluations & Datasets Track). It contains 417,280 calibrated stereo sessions of 128 photorealistic RenderPeople avatars rendered in Unreal Engine 5 across 109 stereo configurations (baseline 5-15 cm, horizontal FOV 10-30 deg, per-camera vergence 0/1/2 deg) and four indoor environments. Each session provides left/right RGB images, metric (cm) and normalized depth maps, surface normals, subject masks, the full OpenCV stereo calibration and rectification matrices, and per-session capture metadata. Splits are identity-disjoint: 103 training subjects (335,780 sessions) and 25 evaluation subjects shared between the validation (70,600 sessions) and test (10,900 sessions) splits. Each record in manifest/sessions.csv corresponds to one stereo capture session; paths are relative to the dataset root and are organized directly under the human subject folders.", |
| "version": "1.0.1", |
| "license": "https://creativecommons.org/licenses/by-nc/4.0/", |
| "url": "https://huggingface.co/datasets/asergiu/HumanStereo", |
| "sameAs": [ |
| "https://github.com/asergiu-ubb/HumanStereoNeurIPS2026Public", |
| "https://huggingface.co/datasets/asergiu/HumanStereoPreview" |
| ], |
| "datePublished": "2026-05-02", |
| "dateModified": "2026-11-03", |
| "citeAs": "@inproceedings{marinescu2026humanstereo, title={HumanStereo: A Benchmark for Metric Facial Depth Estimation and Its Evaluation}, author={Marinescu, Alexandru I. and Borza, Diana L. and D\\u0103r\\u0103bant, Adrian S.}, booktitle={Advances in Neural Information Processing Systems (NeurIPS), Evaluations \\& Datasets Track}, year={2026}}", |
| "keywords": [ |
| "stereo vision", |
| "synthetic dataset", |
| "Unreal Engine 5", |
| "RenderPeople", |
| "depth estimation", |
| "disparity estimation", |
| "human reconstruction", |
| "computer vision", |
| "human faces", |
| "facial depth estimation", |
| "metric depth", |
| "benchmark", |
| "NeurIPS 2026" |
| ], |
| "creator": [ |
| { |
| "@type": "Person", |
| "name": "Alexandru I. Marinescu", |
| "email": "alexandru.marinescu@ubbcluj.ro", |
| "affiliation": { |
| "@type": "Organization", |
| "name": "Babeș-Bolyai University, Department of Computer Science", |
| "url": "https://www.cs.ubbcluj.ro/" |
| } |
| }, |
| { |
| "@type": "Person", |
| "name": "Diana L. Borza", |
| "email": "diana.borza@ubbcluj.ro", |
| "affiliation": { |
| "@type": "Organization", |
| "name": "Babeș-Bolyai University, Department of Computer Science", |
| "url": "https://www.cs.ubbcluj.ro/" |
| } |
| }, |
| { |
| "@type": "Person", |
| "name": "Adrian S. Dărăbant", |
| "email": "sergiu.darabant@ubbcluj.ro", |
| "affiliation": { |
| "@type": "Organization", |
| "name": "Babeș-Bolyai University, Department of Computer Science", |
| "url": "https://www.cs.ubbcluj.ro/" |
| } |
| } |
| ], |
| "publisher": { |
| "@type": "Organization", |
| "name": "Babeș-Bolyai University, Department of Computer Science", |
| "url": "https://www.cs.ubbcluj.ro/" |
| }, |
| "funder": [ |
| { |
| "@type": "Organization", |
| "name": "Romanian Hub for Artificial Intelligence - HRIA (2021-2027, MySMIS no. 351416)" |
| }, |
| { |
| "@type": "Organization", |
| "name": "Project PN-IV-P7-7.1-PTE-2024-0105, contract no. 43PTE (05/05/2025)" |
| } |
| ], |
| "conformsTo": "http://mlcommons.org/croissant/1.1", |
| "inLanguage": "en", |
| "rai:dataCollection": "All data were generated synthetically with an automated Unreal Engine 5 rendering pipeline. For each of the 128 RenderPeople avatars, every valid stereo configuration (109) was rendered in 4 indoor environments; per session the head pose (yaw/pitch/roll in [-10, 10] deg) and the subject distance (20-200 cm) were sampled randomly. Ten capture sessions were rendered per configuration when the vergence angle is 0 deg and five otherwise. No real people were photographed or recorded.", |
| "rai:dataCollectionType": [ |
| "Synthetic data generation (Unreal Engine 5 rendering of licensed 3D scans)" |
| ], |
| "rai:dataCollectionTimeframe": "April 2025 - February 2026", |
| "rai:dataCollectionRawData": "Raw renders are distributed as-is: unrectified 512x512 PNG stereo pairs plus float16/boolean NPZ arrays (depth in centimeters, normalized depth, normals, masks) and the exact OpenCV calibration/rectification matrices (YAML).", |
| "rai:dataCollectionMissingData": "None. Every session contains the complete set of resources; annotations are complete by construction.", |
| "rai:dataPreprocessingProtocol": [ |
| "No filtering or cleaning was applied to the rendered data. Depth maps are stored in metric units (cm). Rectification is not pre-applied; the released evaluation code rectifies on the fly using the stored R1/R2/P1/P2/Q matrices." |
| ], |
| "rai:dataAnnotationProtocol": "Depth maps, surface normals, subject masks, and camera calibration are extracted directly from the rendering engine and are therefore exact (noise-free). Facial landmarks used by the L3DE metric (478-point MediaPipe Face Mesh topology) are detected automatically from the rendered images with the script released in the code repository; no human annotators were involved.", |
| "rai:dataAnnotationPlatform": [ |
| "None (automatic annotations from the rendering engine; MediaPipe Face Mesh for landmarks)" |
| ], |
| "rai:dataAnnotationAnalysis": "Annotations are generated deterministically by the renderer, so inter-annotator agreement does not apply. Known imperfections are rendering artifacts such as aliasing at object boundaries.", |
| "rai:machineAnnotationTools": [ |
| "Unreal Engine 5 (depth, normals, masks, calibration)", |
| "MediaPipe Face Mesh (facial landmarks)" |
| ], |
| "rai:annotationsPerItem": "One complete set of annotations per stereo session (left and right camera).", |
| "rai:annotatorDemographics": "Not applicable (no human annotators).", |
| "rai:personalSensitiveInformation": [ |
| "The dataset contains no real personal data: all images are renderings of commercially licensed 3D scans (RenderPeople). Avatars are labelled with coarse age group, gender and ethnicity to enable fairness analyses. The real optometric captures used for the downstream evaluation in the paper are NOT part of this dataset and are not released." |
| ], |
| "rai:dataUseCases": [ |
| "Benchmarking and training stereo depth estimation for human faces", |
| "Evaluation of metric facial geometry (inter-pupillary distance, landmark distances) for eyewear fitting and AR/VR calibration research", |
| "Surface normal prediction, face segmentation, facial landmark detection, stereo calibration research, synthetic-to-real domain adaptation" |
| ], |
| "rai:dataLimitations": [ |
| "Entirely synthetic: a domain gap with real imagery remains (skin reflectance, hair, eyeglasses, sensor noise, calibration errors are not modelled).", |
| "Avatars are static: neutral expression and rigid head pose only; no dynamic facial expressions or soft-tissue deformation.", |
| "Lighting is not an explicitly controlled variable (it varies only through the four environments).", |
| "Only 128 identities; the evaluation split contains no subject of the Indian ethnicity group (all 6 are in the training split).", |
| "Not a substitute for clinical validation on real patient data." |
| ], |
| "rai:dataBiases": [ |
| "Demographic composition (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. The 19-30 age group and White subjects are over-represented.", |
| "Fronto-parallel rigs (vergence 0 deg) are over-sampled by design (10 vs 5 sessions per configuration) because most pretrained stereo models assume this setup." |
| ], |
| "rai:dataSocialImpact": "Intended to improve metric facial depth estimation for assistive and clinical applications (eyewear fitting, AR/VR interpupillary calibration). Facial depth models could be misused for surveillance or unauthorized biometric identification; the dataset is released under CC BY-NC 4.0 and must not be used for facial recognition, surveillance, or any application that could harm individuals.", |
| "rai:dataReleaseMaintenancePlan": "Hosted on Hugging Face (asergiu/HumanStereo) and maintained by Prof. Adrian-Sergiu Darabant, Babes-Bolyai University (sergiu.darabant@ubbcluj.ro). Errata will be documented on the dataset page; all released versions remain available with version identifiers. Planned future versions will add dynamic facial expressions and controlled lighting.", |
| "distribution": [ |
| { |
| "@type": "cr:FileObject", |
| "@id": "sessions-csv", |
| "name": "sessions.csv", |
| "description": "Unified manifest with one row per stereo capture session. Paths are relative to the dataset root and begin with the human subject folder.", |
| "contentUrl": "manifest/sessions.csv", |
| "encodingFormat": "text/csv", |
| "sha256": "d61ef45d5cd76027995ee6daffa4293b674a817be52bb6fdd028cb809be61a1e" |
| }, |
| { |
| "@type": "cr:FileSet", |
| "@id": "png-images", |
| "name": "PNG RGB images", |
| "encodingFormat": "image/png", |
| "includes": "**/*.png" |
| }, |
| { |
| "@type": "cr:FileSet", |
| "@id": "npz-arrays", |
| "name": "NumPy NPZ array files", |
| "encodingFormat": "application/octet-stream", |
| "includes": "**/*.npz" |
| }, |
| { |
| "@type": "cr:FileSet", |
| "@id": "yaml-calibration", |
| "name": "OpenCV YAML calibration files", |
| "encodingFormat": "application/x-yaml", |
| "includes": "**/*.yml" |
| }, |
| { |
| "@type": "cr:FileSet", |
| "@id": "json-metadata", |
| "name": "Per-session JSON metadata files", |
| "encodingFormat": "application/json", |
| "includes": "**/*.json" |
| } |
| ], |
| "recordSet": [ |
| { |
| "@type": "cr:RecordSet", |
| "@id": "stereo_sessions", |
| "name": "stereo_sessions", |
| "description": "One record per stereo capture session.", |
| "key": { |
| "@id": "stereo_sessions/session_id" |
| }, |
| "field": [ |
| { |
| "@type": "cr:Field", |
| "@id": "stereo_sessions/session_id", |
| "name": "session_id", |
| "description": "Unique stereo session identifier.", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "sessions-csv" |
| }, |
| "extract": { |
| "column": "session_id" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "stereo_sessions/split", |
| "name": "split", |
| "description": "Dataset split: train, validation, or test.", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "sessions-csv" |
| }, |
| "extract": { |
| "column": "split" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "stereo_sessions/apartment_location", |
| "name": "apartment_location", |
| "description": "Apartment location code: 1=living room, 2=bedroom, 3=kitchen, 4=bathroom.", |
| "dataType": "sc:Integer", |
| "source": { |
| "fileObject": { |
| "@id": "sessions-csv" |
| }, |
| "extract": { |
| "column": "apartment_location" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "stereo_sessions/apartment_location_name", |
| "name": "apartment_location_name", |
| "description": "Human-readable apartment location name derived from the location code.", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "sessions-csv" |
| }, |
| "extract": { |
| "column": "apartment_location_name" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "stereo_sessions/human_subject_name", |
| "name": "human_subject_name", |
| "description": "RenderPeople human subject/avatar name.", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "sessions-csv" |
| }, |
| "extract": { |
| "column": "human_subject_name" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "stereo_sessions/stereo_setup_index", |
| "name": "stereo_setup_index", |
| "description": "Stereo setup index.", |
| "dataType": "sc:Integer", |
| "source": { |
| "fileObject": { |
| "@id": "sessions-csv" |
| }, |
| "extract": { |
| "column": "stereo_setup_index" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "stereo_sessions/capture_session_index", |
| "name": "capture_session_index", |
| "description": "Capture session index within the stereo setup.", |
| "dataType": "sc:Integer", |
| "source": { |
| "fileObject": { |
| "@id": "sessions-csv" |
| }, |
| "extract": { |
| "column": "capture_session_index" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "stereo_sessions/session_dir", |
| "name": "session_dir", |
| "description": "Relative directory containing this session resources; top-level split folder has been removed, so paths begin with the human subject folder.", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "sessions-csv" |
| }, |
| "extract": { |
| "column": "session_dir" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "stereo_sessions/image_left", |
| "name": "image_left", |
| "description": "Relative path to left RGB image file, PNG.", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "sessions-csv" |
| }, |
| "extract": { |
| "column": "image_left" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "stereo_sessions/image_right", |
| "name": "image_right", |
| "description": "Relative path to right RGB image file, PNG.", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "sessions-csv" |
| }, |
| "extract": { |
| "column": "image_right" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "stereo_sessions/mask_left", |
| "name": "mask_left", |
| "description": "Relative path to left avatar/background boolean mask, compressed NumPy NPZ, expected shape 512x512.", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "sessions-csv" |
| }, |
| "extract": { |
| "column": "mask_left" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "stereo_sessions/mask_right", |
| "name": "mask_right", |
| "description": "Relative path to right avatar/background boolean mask, compressed NumPy NPZ, expected shape 512x512.", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "sessions-csv" |
| }, |
| "extract": { |
| "column": "mask_right" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "stereo_sessions/normal_map_left", |
| "name": "normal_map_left", |
| "description": "Relative path to left world-space normal map, compressed NumPy NPZ, expected shape 512x512x3.", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "sessions-csv" |
| }, |
| "extract": { |
| "column": "normal_map_left" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "stereo_sessions/normal_map_right", |
| "name": "normal_map_right", |
| "description": "Relative path to right world-space normal map, compressed NumPy NPZ, expected shape 512x512x3.", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "sessions-csv" |
| }, |
| "extract": { |
| "column": "normal_map_right" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "stereo_sessions/depth_map_norm_left", |
| "name": "depth_map_norm_left", |
| "description": "Relative path to left normalized depth map, compressed NumPy NPZ, expected shape 512x512, values in [0, 1].", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "sessions-csv" |
| }, |
| "extract": { |
| "column": "depth_map_norm_left" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "stereo_sessions/depth_map_norm_right", |
| "name": "depth_map_norm_right", |
| "description": "Relative path to right normalized depth map, compressed NumPy NPZ, expected shape 512x512, values in [0, 1].", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "sessions-csv" |
| }, |
| "extract": { |
| "column": "depth_map_norm_right" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "stereo_sessions/depth_map_phys_left", |
| "name": "depth_map_phys_left", |
| "description": "Relative path to left physical depth map, compressed NumPy NPZ, expected shape 512x512, units in centimeters.", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "sessions-csv" |
| }, |
| "extract": { |
| "column": "depth_map_phys_left" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "stereo_sessions/depth_map_phys_right", |
| "name": "depth_map_phys_right", |
| "description": "Relative path to right physical depth map, compressed NumPy NPZ, expected shape 512x512, units in centimeters.", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "sessions-csv" |
| }, |
| "extract": { |
| "column": "depth_map_phys_right" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "stereo_sessions/calibration", |
| "name": "calibration", |
| "description": "Relative path to OpenCV YAML stereo calibration and rectification file for the left/right stereo pair.", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "sessions-csv" |
| }, |
| "extract": { |
| "column": "calibration" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "stereo_sessions/metadata", |
| "name": "metadata", |
| "description": "Relative path to per-session JSON metadata containing camera baseline, vergence, field of view, and avatar distance.", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "sessions-csv" |
| }, |
| "extract": { |
| "column": "metadata" |
| } |
| } |
| } |
| ] |
| } |
| ] |
| } |
