| { |
| "@context": { |
| "@language": "en", |
| "@vocab": "https://schema.org/", |
| "citeAs": "cr:citeAs", |
| "column": "cr:column", |
| "conformsTo": "dct:conformsTo", |
| "cr": "http://mlcommons.org/croissant/", |
| "rai": "http://mlcommons.org/croissant/RAI/", |
| "data": { |
| "@id": "cr:data", |
| "@type": "@json" |
| }, |
| "dataType": { |
| "@id": "cr:dataType", |
| "@type": "@vocab" |
| }, |
| "dct": "http://purl.org/dc/terms/", |
| "examples": { |
| "@id": "cr:examples", |
| "@type": "@json" |
| }, |
| "extract": "cr:extract", |
| "field": "cr:field", |
| "fileProperty": "cr:fileProperty", |
| "fileObject": "cr:fileObject", |
| "fileSet": "cr:fileSet", |
| "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", |
| "samplingRate": "cr:samplingRate", |
| "sc": "https://schema.org/", |
| "separator": "cr:separator", |
| "source": "cr:source", |
| "subField": "cr:subField", |
| "transform": "cr:transform" |
| }, |
| "@type": "sc:Dataset", |
| "conformsTo": "http://mlcommons.org/croissant/1.0", |
| "name": "CM-EVS", |
| "description": "CM-EVS is a curated panoramic RGB-D dataset built under a single principle: maximize the geometric coverage of a 3D scene with the fewest equirectangular (ERP) frames possible. The headline release contains 11,583 ERP RGB-depth-pose frames over 326 Blender indoor scenes (CC-BY 4.0), each paired with the per-step provenance log of the depth-conflict-aware curator that selected it. The full v1.0 release additionally provides 786,344 frames re-encoded from TartanGround (783,944 frames over 63 environments) and OB3D (2,400 frames over 12 scenes) outdoor sources into the same ERP and world-to-camera pose schema, plus license-aware adapter packages for HM3D (14,475 frames over 401 rooms after local regeneration) and ScanNet++ (8,267 frames over 500 scans after local regeneration) that produce matched frames locally without redistributing licensed assets.", |
| "version": "1.0.0", |
| "license": "https://creativecommons.org/licenses/by/4.0/", |
| "url": "https://huggingface.co/datasets/anon-cmevs-2026/cmevs-erp-eval", |
| "citeAs": "@inproceedings{cmevs2026, title={{CM-EVS}: A Coverage-Curated Panoramic {RGB-D} Dataset for Indoor Scene Understanding}, author={Anonymous Author(s)}, booktitle={NeurIPS 2026 Datasets and Benchmarks Track (under review)}, year={2026}}", |
| "creator": { |
| "@type": "Organization", |
| "name": "Anonymous (double-blind submission)" |
| }, |
| "datePublished": "2026-05-01", |
| "keywords": [ |
| "panoramic", |
| "equirectangular", |
| "ERP", |
| "RGB-D", |
| "view planning", |
| "fixed-budget", |
| "data-centric", |
| "viewpoint provenance", |
| "indoor scene understanding", |
| "panoramic depth estimation", |
| "novel view synthesis", |
| "world model pretraining" |
| ], |
| "isLiveDataset": false, |
| "distribution": [ |
| { |
| "@type": "cr:FileObject", |
| "@id": "blender-indoor-archive.tar", |
| "name": "blender-indoor-archive.tar", |
| "contentUrl": "https://huggingface.co/datasets/anon-cmevs-2026/cmevs-erp-eval/resolve/main/blender_indoor.tar", |
| "encodingFormat": "application/x-tar", |
| "sha256": "TODO_SHA256" |
| }, |
| { |
| "@type": "cr:FileSet", |
| "@id": "blender-indoor-rgb", |
| "name": "blender-indoor-rgb", |
| "containedIn": { |
| "@id": "blender-indoor-archive.tar" |
| }, |
| "encodingFormat": "image/png", |
| "includes": "rgb/*.png" |
| }, |
| { |
| "@type": "cr:FileSet", |
| "@id": "blender-indoor-depth", |
| "name": "blender-indoor-depth", |
| "containedIn": { |
| "@id": "blender-indoor-archive.tar" |
| }, |
| "encodingFormat": "application/octet-stream", |
| "includes": "depth/*.npy" |
| }, |
| { |
| "@type": "cr:FileSet", |
| "@id": "blender-indoor-pose", |
| "name": "blender-indoor-pose", |
| "containedIn": { |
| "@id": "blender-indoor-archive.tar" |
| }, |
| "encodingFormat": "application/json", |
| "includes": "pose/*.json" |
| }, |
| { |
| "@type": "cr:FileSet", |
| "@id": "blender-indoor-metadata", |
| "name": "blender-indoor-metadata", |
| "containedIn": { |
| "@id": "blender-indoor-archive.tar" |
| }, |
| "encodingFormat": "application/json", |
| "includes": "metadata/*.json*" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "outdoor-tartanground-adapter.tar", |
| "name": "outdoor-tartanground-adapter.tar", |
| "contentUrl": "https://huggingface.co/datasets/anon-cmevs-2026/cmevs-erp-eval/resolve/main/outdoor_tartanground_adapter.tar", |
| "encodingFormat": "application/x-tar", |
| "sha256": "TODO_SHA256" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "outdoor-ob3d-adapter.tar", |
| "name": "outdoor-ob3d-adapter.tar", |
| "contentUrl": "https://huggingface.co/datasets/anon-cmevs-2026/cmevs-erp-eval/resolve/main/outdoor_ob3d_adapter.tar", |
| "encodingFormat": "application/x-tar", |
| "sha256": "TODO_SHA256" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "hm3d-adapter.tar", |
| "name": "hm3d-adapter.tar", |
| "contentUrl": "https://huggingface.co/datasets/anon-cmevs-2026/cmevs-erp-eval/resolve/main/hm3d_adapter.tar", |
| "encodingFormat": "application/x-tar", |
| "sha256": "TODO_SHA256" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "scannetpp-adapter.tar", |
| "name": "scannetpp-adapter.tar", |
| "contentUrl": "https://huggingface.co/datasets/anon-cmevs-2026/cmevs-erp-eval/resolve/main/scannetpp_adapter.tar", |
| "encodingFormat": "application/x-tar", |
| "sha256": "TODO_SHA256" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "curator-source-code.tar", |
| "name": "curator-source-code.tar", |
| "contentUrl": "https://huggingface.co/datasets/anon-cmevs-2026/cmevs-erp-eval/resolve/main/code.tar", |
| "encodingFormat": "application/x-tar", |
| "sha256": "TODO_SHA256" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "documentation.tar", |
| "name": "documentation.tar", |
| "contentUrl": "https://huggingface.co/datasets/anon-cmevs-2026/cmevs-erp-eval/resolve/main/docs.tar", |
| "encodingFormat": "application/x-tar", |
| "sha256": "TODO_SHA256" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "frame-manifest.csv", |
| "name": "frame-manifest.csv", |
| "contentUrl": "https://huggingface.co/datasets/anon-cmevs-2026/cmevs-erp-eval/resolve/main/frame_manifest.csv", |
| "encodingFormat": "text/csv", |
| "sha256": "TODO_SHA256" |
| } |
| ], |
| "recordSet": [ |
| { |
| "@type": "cr:RecordSet", |
| "@id": "erp-frame-records", |
| "name": "erp-frame-records", |
| "description": "One record per released ERP frame. Curator-only fields (viewpoint_score, coverage_gain, conflict_ratio, candidate_id) are populated only for frames produced by the depth-conflict-aware curator; outdoor re-encoded frames carry the schema fields without per-step provenance.", |
| "field": [ |
| { |
| "@type": "cr:Field", |
| "@id": "erp-frame-records/frame_id", |
| "name": "frame_id", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "frame-manifest.csv" |
| }, |
| "extract": { |
| "column": "frame_id" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "erp-frame-records/source", |
| "name": "source", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "frame-manifest.csv" |
| }, |
| "extract": { |
| "column": "source" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "erp-frame-records/scene_id", |
| "name": "scene_id", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "frame-manifest.csv" |
| }, |
| "extract": { |
| "column": "scene_id" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "erp-frame-records/room_id", |
| "name": "room_id", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "frame-manifest.csv" |
| }, |
| "extract": { |
| "column": "room_id" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "erp-frame-records/split", |
| "name": "split", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "frame-manifest.csv" |
| }, |
| "extract": { |
| "column": "split" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "erp-frame-records/rgb", |
| "name": "rgb", |
| "dataType": "sc:ImageObject", |
| "source": { |
| "fileObject": { |
| "@id": "frame-manifest.csv" |
| }, |
| "extract": { |
| "column": "rgb_path" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "erp-frame-records/depth", |
| "name": "depth", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "frame-manifest.csv" |
| }, |
| "extract": { |
| "column": "depth_path" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "erp-frame-records/pose_quaternion", |
| "name": "pose_quaternion", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "frame-manifest.csv" |
| }, |
| "extract": { |
| "column": "pose_quaternion" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "erp-frame-records/pose_position", |
| "name": "pose_position", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "frame-manifest.csv" |
| }, |
| "extract": { |
| "column": "pose_position" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "erp-frame-records/camera_type", |
| "name": "camera_type", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "frame-manifest.csv" |
| }, |
| "extract": { |
| "column": "camera_type" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "erp-frame-records/viewpoint_score", |
| "name": "viewpoint_score", |
| "dataType": "sc:Float", |
| "source": { |
| "fileObject": { |
| "@id": "frame-manifest.csv" |
| }, |
| "extract": { |
| "column": "viewpoint_score" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "erp-frame-records/coverage_gain", |
| "name": "coverage_gain", |
| "dataType": "sc:Float", |
| "source": { |
| "fileObject": { |
| "@id": "frame-manifest.csv" |
| }, |
| "extract": { |
| "column": "coverage_gain" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "erp-frame-records/conflict_ratio", |
| "name": "conflict_ratio", |
| "dataType": "sc:Float", |
| "source": { |
| "fileObject": { |
| "@id": "frame-manifest.csv" |
| }, |
| "extract": { |
| "column": "conflict_ratio" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "erp-frame-records/candidate_id", |
| "name": "candidate_id", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "frame-manifest.csv" |
| }, |
| "extract": { |
| "column": "candidate_id" |
| } |
| } |
| } |
| ] |
| } |
| ], |
| "rai:dataCollection": "Indoor data is produced by the CM-EVS pipeline (asset loading, coordinate normalization, candidate generation, 26-direction geometric-validity filtering, conflict-aware greedy selection, 2048x1024 high-resolution Cycles ERP rendering, export under the unified schema). Outdoor data is sourced from TartanGround and OB3D and re-encoded into the unified schema; the curator is not run on outdoor sources in v1.0. HM3D and ScanNet++ frames are not redistributed; the release ships adapter regeneration scripts.", |
| "rai:dataPreprocessingProtocol": "Coordinate normalization to a right-handed +X-right, +Y-up, +Z-forward world frame with the OpenCV-style camera frame; pose stored as a scalar-first world-to-camera quaternion plus a position relative to the scene's first selected frame. AABB computation; source-specific candidate generation; 26-direction geometric-validity filter. Cubemap-to-ERP re-encoding at native resolution for outdoor sources; optional exposure adjustment for Blender; output schema conversion. Candidate probes, intermediate caches, pre-render-all oracle frames, and locally regenerated HM3D / ScanNet++ outputs are excluded from the public frame count F_pub.", |
| "rai:dataAnnotationProtocol": "No human annotation is performed. All labels (split, source, scene id, viewpoint score, coverage gain, conflict ratio) are produced automatically by the curator pipeline and recorded in metadata/per_step_log.jsonl and metadata/selected_viewpoints.json.", |
| "rai:dataReleaseMaintenancePlan": "Versioned releases on a 6-month cadence. Errata tracked via the project repository; SHA256 manifests refreshed at every release; HM3D and ScanNet++ regeneration scripts updated when upstream APIs, file layouts, or access terms change.", |
| "rai:dataUseCases": [ |
| "Panoramic depth estimation", |
| "ERP novel-view synthesis", |
| "Panoramic Gaussian-splatting reconstruction", |
| "Panoramic world-model pretraining", |
| "Fixed-budget viewpoint policy evaluation" |
| ], |
| "rai:dataLimitations": [ |
| "Real-scan derived frames (HM3D, ScanNet++) are not redistributed; users must accept upstream license terms and regenerate locally.", |
| "Outdoor frames are re-encoded source trajectories rather than curator-selected subsets and therefore do not carry per-step provenance.", |
| "Synthetic-real transfer must be validated separately by source; we do not claim Blender-only gains imply real-scan gains.", |
| "Geometry-validity filters may fail in atria, semi-outdoor spaces, narrow transitions, noisy scans, or pure point-cloud scenes." |
| ], |
| "rai:personalSensitiveInformation": "No new personal data is collected. Real-scan sources (HM3D, ScanNet++) may depict private indoor layouts and are not redistributed as derived frames. Even regeneration scripts and viewpoint metadata can reveal where observations would be sampled within a private space; users must comply with upstream source access terms.", |
| "rai:dataBiases": [ |
| "Source assets inherit geographic, architectural, and scanning biases.", |
| "HM3D and ScanNet++ are skewed toward scanned residential indoor spaces.", |
| "Blender assets are skewed toward staged residential, office, and architectural scenes.", |
| "Outdoor sources (TartanGround, OB3D) are skewed toward simulator-generated terrain along circular trajectories.", |
| "Synthetic Blender materials may not match real-scan sensor noise." |
| ], |
| "rai:dataSocialImpact": "CM-EVS lowers the engineering cost of producing auditable panoramic RGB-D resources from existing 3D scenes. Positive uses include panoramic perception, data-centric evaluation, view-planning research, and 3D-consistent world-model pretraining. Potential harms include over-trusting synthetic data, obscuring upstream dataset bias, and using real indoor scans in privacy-sensitive settings. The release therefore separates public synthetic frames from licensed real-scan regeneration and documents intended uses, non-uses, and source licenses." |
| } |
