ManfredC/cmevs-erp-eval

CM-EVS: A Coverage-Curated Panoramic RGB-D Dataset for Indoor Scene Understanding

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 release is structured as one redistributable Blender indoor data archive plus four license-aware adapter packages that regenerate matched frames locally from upstream sources whose terms forbid redistribution.

v1.0 status: this version stages the full Blender indoor data drop (374 scene instances, 13,631 ERP RGB-depth-pose frames; 201 from the round1+2 sampling and 173 from round2). The paper's headline 326 scenes / 11,583 frames is the curator-selected subset that will be derived from this drop after the §5 evaluation experiments finalize. Result CSVs under results/ are schema stubs only in this release (see results/README.md).

Dataset summary

Source	License	Released here	Scenes / frames
Blender indoor	CC-BY 4.0	Full data (blender_indoor/)	374 / 13,631
HM3D	upstream EULA	Adapter only (adapters/hm3d/)	401 rooms / regen-only
ScanNet++	upstream ToS	Adapter only (adapters/scannetpp/)	500 scans / regen-only
OB3D (outdoor)	upstream license	Adapter only (adapters/ob3d/)	24 / regen-only
TartanGround (outdoor)	upstream license	Adapter only (adapters/tartanground/)	762 parts / regen-only

The Blender indoor frames are the only redistributable RGB-D data. For the four restricted sources, this dataset ships the per-source adapter (config + pipeline script + scene-id metadata); users obtain upstream data themselves and run the adapter locally to reproduce matching ERP frames under the unified schema.

Output schema

Every released ERP frame follows a single coordinate convention:

• World frame: right-handed, +X right, +Y up, +Z forward
• Camera frame: OpenCV (+x image right, +y image down, +z camera forward)
• Pose: scalar-first quaternion q_wc = [w, x, y, z] plus position C_w − C_{w,0} relative to the scene's first selected frame (the absolute first-frame center is recorded once per scene in meta.json when present)
• ERP pixel coords: longitude (u/W − 0.5) · 2π, latitude (0.5 − v/H) · π
• Range depth: each pixel stores radial distance from camera center to surface (not perspective z-depth). NaN or 0 marks invalid pixels.

File	Format	Description
panorama_{NNNN}.png	PNG, 2048×1024	ERP RGB image
panorama_{NNNN}_depth.npy	float32 array	ERP range depth (m); NaN or 0 if invalid; absent for some frames where depth was not produced
pose_{NNNN}.json	JSON	q_wc, position, camera_type

Per-scene meta.json, metadata/selected_viewpoints.json, metadata/candidates.jsonl, metadata/per_step_log.jsonl (curator-only) are not yet present for the full merged 374-scene set; they will be added in a later release once the curator run completes.

Directory layout

cmevs_hf_release/
├── README.md      (this file — HF dataset card)
├── LICENSE.md      (mixed-license matrix)
├── CHANGELOG.md
├── croissant.json    (MLCommons Croissant v1.0; passes mlcroissant 1.1 validator)
├── SHA256SUMS      (top-level checksums, excluding blender_indoor/scenes/)
├── blender_indoor/
│  ├── README.md
│  ├── scenes/sence_indoor_{0001..0374}/{panorama,pose}_{NNNN}.{png,npy,json}
│  ├── SHA256SUMS    (40,893 lines for 13,631 RGB-depth-pose files)
│  └── metadata/{source_manifest.json, splits.json, frame_manifest.csv,
│         scene_id_mapping.csv, frame_id_mapping.csv}
├── adapters/{hm3d, scannetpp, ob3d, tartanground}/
│  ├── README.md
│  ├── config.yaml
│  ├── pipeline.py / reencoding_script.md
│  └── metadata/source_manifest.json
├── code/        (curator core modules + scripts; reviewer reference)
└── results/       (paper §5 CSV schemas; stubs only — no metrics yet; see results/README.md)

Datasheet

Following Gebru et al. 2021. (Source: main.tex Appendix A — content here is a faithful markdown rendering; cross-check against the paper PDF.)

Motivation

Purpose. Evaluates fixed-budget panoramic viewpoint curation policies on existing 3D assets, and provides reproducible ERP RGB-D-pose samples for panoramic perception experiments.

Creators / funding. Creators: Jiale Liu*, Jungang Li*, Jieming Yu*, Xinglin Yu*, Zihao Dongfang, Keyu Shi, Zongjian Ding, Jiahuan Zhang, Shunwen Bai, Haoran Huang, Yurun Wang, Yanxi Wu, Ningzhe Yu, Yudong Gao†, Mingjun Cheng† (* equal contribution; † corresponding authors: Yudong Gao, Mingjun Cheng). Affiliations: Zhejiang University (Jiale Liu, Shunwen Bai, Haoran Huang, Yurun Wang, Yanxi Wu, Ningzhe Yu, Mingjun Cheng); The Hong Kong University of Science and Technology (Guangzhou) (Jungang Li, Zihao Dongfang); The Hong Kong University of Science and Technology (Jieming Yu, Yudong Gao); Beijing Institute of Technology, Zhuhai (Xinglin Yu); Harbin Institute of Technology (Keyu Shi); University of Chinese Academy of Sciences (Zongjian Ding); Tianjin University (Jiahuan Zhang).

Composition

Instances. Each instance is an ERP frame triple (RGB image + range-depth array + camera pose), plus per-scene meta.json and curator-only provenance metadata.

Counts. v1.0 stages 13,631 ERP RGB-depth-pose frames across 374 Blender indoor scene instances (CC-BY 4.0). A 2026-06-01 quality replacement refreshed 122 Blender indoor scene directories while preserving the release-level scene/frame totals; RGB, depth, and pose counts are now all 13,631. Provenance for replaced scene contents is recorded in blender_indoor/metadata/scene_id_mapping.csv and frame_id_mapping.csv. The four restricted sources (HM3D / ScanNet++ / OB3D / TartanGround) ship adapters only; users regenerate matching frames locally.

Sampling. Indoor (Blender) frames are produced offline by Cycles ERP rendering. The 374 scene instances comprise 201 from round1+2 (Blender_indoor_FOU_threshold-0.2, rounds 1+2 merge) and 173 from round2 (independent extraction). 48 original sence_indoor_XXXX ids appear in both rounds with different sampling outcomes; both versions are kept and renumbered (see blender_indoor/metadata/scene_id_mapping.csv for traceability). Outdoor source trajectories (TartanGround, OB3D) are re-encoded into the unified schema by the adapters/{tartanground,ob3d}/ pipelines; the curator does not run on outdoor sources in v1.0.

Fields. RGB PNG (2048×1024 for Blender indoor; native source resolution otherwise), float32 range depth (.npy), pose JSON with scalar-first q_wc, meta.json, candidate / viewpoint / per-step-log metadata (curator-produced frames only), source / scene / split ids.

Missing values. Invalid depth pixels are NaN or 0 by source convention; per-frame invalid-depth ratio statistics are planned for results/frame_quality.csv (schema stub only in this release; see results/README.md).

Splits. Default scene-level 70 / 15 / 15 split via sha256(new_scene_id) % 100. See blender_indoor/metadata/splits.json. The downstream panoramic-depth experiment (paper §4.10) uses a separate 94-scene Blender-indoor subset under its own scene-level split (84 / 10 / 10 = 3,400 / 362 / 423 frames).

Collection

Indoor data is produced by the CM-EVS pipeline (asset loading, coordinate normalization, candidate generation, 26-direction geometric-validity filtering, conflict-aware greedy selection, 2048×1024 high-resolution Cycles ERP rendering, export under the unified schema). Outdoor data (TartanGround, OB3D) is re-encoded into the unified schema; the curator does not run on outdoor sources in v1.0. HM3D and ScanNet++ frames are not redistributed: the release ships adapter regeneration scripts that produce matched frames locally after the user accepts upstream license terms. No new human-subject data is collected.

Preprocessing

Coordinate normalization to right-handed +X-right +Y-up +Z-forward world frame with the OpenCV-style camera frame; pose stored as scalar-first q_wc = [w, x, y, z] plus 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. Raw upstream 3D assets are not redistributed unless upstream licenses allow.

Uses

Recommended: panoramic depth estimation, ERP novel-view synthesis, data-centric viewpoint policy comparison, view-planning research, panoramic Gaussian-splatting reconstruction, panoramic world-model pretraining.

Avoid: identity-sensitive inference, safety-critical deployment, claims about private indoor spaces, treating synthetic-only results as real-world evidence without further validation.

Distribution

Blender indoor frames (CC-BY 4.0), curator code (MIT), documentation (CC-BY 4.0), Datasheet, and Croissant metadata are released here. The four restricted sources ship metadata + regeneration scripts only.

Maintenance

Versioned releases on a 6-month cadence. Errata are tracked via the project repository; checksum manifests are refreshed at every release; regeneration scripts are updated when upstream APIs, file layouts, or access terms change.

Code

The full curator source code, adapters, and reproduction scripts are released as a separate public repository:

Code: huggingface.co/ManfredC/cmevs-code

The code/ subtree mirrored inside this dataset repository is provided for offline reviewer convenience; the linked code repository is the canonical source.

Citation

@inproceedings{cmevs2026,
 title={{CM-EVS}: A Coverage-Curated Panoramic {RGB-D} Dataset for Indoor Scene Understanding},
 author={Jiale Liu and Jungang Li and Jieming Yu and Xinglin Yu and Zihao Dongfang and Keyu Shi and Zongjian Ding and Jiahuan Zhang and Shunwen Bai and Haoran Huang and Yurun Wang and Yanxi Wu and Ningzhe Yu and Yudong Gao and Mingjun Cheng},
 booktitle={Advances in Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track},
 year={2026}
}

Reviewer quick sample (NeurIPS 2026 D&B "large dataset URL" requirement)

The full Blender indoor archive is ~109 GB. To support reviewer-time inspection without a full download, scene sence_indoor_0001 is provided as a representative sample at:

huggingface.co/datasets/ManfredC/cmevs-erp-eval/tree/main/blender_indoor/scenes/sence_indoor_0001

This single scene contains 33 RGB panoramas (panorama_NNNN.png, 2048×1024 ERP), 33 range-depth arrays (panorama_NNNN_depth.npy, float32 metres), and 33 pose JSON files (pose_NNNN.json, scalar-first quaternion + position) — 99 files in total, ~250 MB.

Sampling methodology

The scene was produced by the same end-to-end CM-EVS pipeline as every other Blender indoor scene in this release: asset loading → coordinate normalization to right-handed +X/+Y/+Z world frame → grid-based candidate generation with the 26-direction geometric-validity filter → conflict-aware greedy viewpoint selection → 2048×1024 Cycles ERP rendering → unified-schema export. No special preprocessing distinguishes the sample from the rest of the release; it was selected only because (i) it is the first scene id in lexical order and (ii) it represents the round1+2 sampling subset (the 201-scene half of the 374-scene v1.0 release; the other 173 scenes come from the round2 independent extraction). Reviewers can therefore use this scene to verify file format, coordinate convention, depth validity statistics, and image-depth alignment for the full release. The complete provenance — including the per-step coverage gain $G_t$, conflict ratio $L_t$, and viewpoint score $s_t$ — is in metadata/per_step_log.jsonl (curator-only fields, populated for all curator-produced frames once the §5 evaluation experiments are finalized).

Notes on directory naming

Scene directories under blender_indoor/scenes/ use the legacy id pattern sence_indoor_NNNN (note: sence, not scene). This is a typo inherited from the upstream Blender source pipeline used to produce the v1.0 build, and it is preserved verbatim so that scene ids match the production-side run logs and the entries in metadata/scene_id_mapping.csv. The misspelling does not affect file content, ERP coordinate convention, depth validity, pose schema, frame indexing, or downstream parsing — only the directory name string. Directories will be renamed to scene_indoor_NNNN in v1.1; the rename will be reflected in a new scene_id_mapping.csv row pointing each new id to its v1.0 sence_indoor_NNNN predecessor so existing consumers continue to resolve.

Verifying integrity

# top-level files + adapter packages + code + metadata
shasum -a 256 -c SHA256SUMS

# Blender indoor frames (40,893 entries: 13,631 panorama + 13,631 depth + 13,631 pose)
cd blender_indoor && shasum -a 256 -c SHA256SUMS

License

See LICENSE.md for the per-component license matrix.