PBR-Rooms: A PBR-Complete Indoor Benchmark for Inverse Rendering Evaluation

Dataset Summary

PBR-Rooms is a synthetic indoor benchmark for evaluating multi-channel inverse rendering under physically based rendering (PBR) ground truth. It is designed to test whether models can recover both scene geometry and PBR material properties from RGB images.

The complete generated PBR-Rooms corpus contains 1,000 procedurally generated indoor scenes across five room categories: bathroom, bedroom, dining room, kitchen, and living room. The complete corpus contains 39,000 rendered RGB images at 1280 × 720 resolution, including 24,000 viewpoint-variation images and 15,000 lighting-variation images.

For anonymous review, this repository is organized as a v0.1 review release. The review release is centered on the official benchmark evaluation subset used in the paper, containing 1,410 rendered RGB images with pixel-aligned ground-truth channels: depth, surface normal, albedo, roughness, metallic, and object segmentation. All benchmark results reported in the submission are computed on this released benchmark subset.

In addition to the official benchmark subset and review sample, this repository may contain category-level tar shards from the complete generated corpus. If not all corpus shards are present, these uploaded tar files should be interpreted as a partial corpus release rather than the complete 39,000-image release. The complete generated corpus is being packaged as a versioned full release, without changing the official benchmark subset, split file, or evaluation protocol used for the submitted results.

PBR-Rooms is intended primarily as an evaluation dataset for comparing dedicated inverse rendering models with general-purpose vision-language or image-generation models. It supports controlled analysis under changes in camera viewpoint and illumination.

Release Status

Scope	Scenes	Images	Status	Purpose
Complete generated corpus	1,000	39,000	Generated; being packaged as a versioned full release	Full corpus analysis and future public release
Official benchmark subset	65	1,410	Released for anonymous review	Reproduce all benchmark results reported in the submission
Review sample	Small subset	Small subset	Released for quick inspection	Inspect file structure and channel examples
Category-level tar shards	Partial corpus shards if not all shards are present	Depends on uploaded shards	Partial corpus release	Additional large-scale inspection and download

The official benchmark subset is the reproducibility target for the submitted benchmark results. Additional corpus shards may be uploaded as part of the versioned full release, but they do not change the benchmark subset, split definition, metrics, or reported evaluation protocol.

Dataset Overview

Property	Complete generated corpus	v0.1 anonymous review release
Dataset type	Synthetic indoor benchmark dataset	Synthetic indoor benchmark artifact
Scenes	1,000	65 benchmark scenes
Rendered RGB images	39,000	1,410 benchmark images
Resolution	1280 × 720	1280 × 720
Room categories	5	5
Channels	RGB, depth, surface normal, albedo, roughness, metallic, object segmentation	Same channels for benchmark subset
Main use	Full corpus analysis	Reproducing submitted benchmark results
License	CC BY-NC 4.0	CC BY-NC 4.0

Intended Use and Evaluation Role

PBR-Rooms is designed for controlled evaluation of inverse rendering models in synthetic indoor scenes. It supports RGB-to-depth, RGB-to-normal, RGB-to-albedo, RGB-to-roughness, RGB-to-metallic, and multi-channel inverse rendering evaluation. The released object segmentation maps are auxiliary annotations for inspection and ablation rather than primary benchmark targets.

The dataset supports the following evaluation claims:

• whether models can recover geometry and PBR material properties from RGB images;
• whether general-purpose vision-language or image-generation models can produce physically meaningful inverse-rendering outputs;
• whether predictions remain robust under controlled viewpoint and illumination variations;
• whether models can predict non-trivial metallic regions rather than producing near-empty or visually ambiguous metallic maps.

PBR-Rooms should not be used as sole evidence for real-world deployment performance. It is synthetic-only and does not fully capture real camera noise, lens artifacts, motion blur, exposure changes, compression artifacts, or the full diversity of real indoor environments. It is not intended for human identification, surveillance, biometric inference, real-estate profiling, privacy-sensitive indoor reconstruction, autonomous driving, or outdoor scene understanding.

Dataset Composition and Statistics

The statistics in this section describe the complete generated corpus unless otherwise specified. The anonymous review release focuses on the official benchmark subset used for the submitted experiments.

PBR-Rooms is balanced across five room categories. Each category contains 200 valid scenes and 7,800 rendered images in the complete generated corpus.

Scene category	Scenes	Images	Viewpoint-varied images	Lighting-varied images
Bathroom	200	7,800	4,800	3,000
Bedroom	200	7,800	4,800	3,000
Dining room	200	7,800	4,800	3,000
Kitchen	200	7,800	4,800	3,000
Living room	200	7,800	4,800	3,000
Total	1,000	39,000	24,000	15,000

Modalities in the Benchmark Subset

Modality	Description	Availability in benchmark subset
RGB	Rendered indoor image	100%
Depth	Metric depth in meters	100%
Surface normal	Camera-space surface normal	100%
Albedo	Intrinsic base color / diffuse reflectance	100%
Roughness	PBR roughness in [0, 1]	100%
Metallic	PBR metallic value in [0, 1]	100%
ObjectSegmentation	Object/semantic segmentation map	100%

Metallic Statistics

Metallic statistics are computed over evaluable pixels. Metallic pixel ratio is the fraction of evaluable pixels whose metallic value exceeds a threshold. Image-level metallic coverage is the fraction of evaluable pixels in an image with metallic > 0.5.

Statistic	Value
Mean metallic value	0.0229
Metallic > 0.05 pixel ratio	4.94%
Metallic > 0.10 pixel ratio	4.81%
Metallic > 0.50 pixel ratio	0.83%
Images with metallic coverage > 0.1%	41.77%
Images with metallic coverage > 1%	12.44%
Images with metallic coverage > 5%	4.04%
Images with metallic coverage > 10%	2.02%
Scenes with at least one view with metallic coverage > 1%	33.57%
Scenes with at least one view with metallic coverage > 5%	15.58%

Scene category	Metallic > 0.05 pixels	Metallic > 0.50 pixels	Images with metallic coverage > 1%	Images with metallic coverage > 5%
Bathroom	6.49%	0.50%	7.47%	3.03%
Bedroom	1.74%	0.19%	2.68%	0.15%
Dining room	1.87%	0.37%	7.69%	0.91%
Kitchen	11.12%	2.50%	33.31%	14.43%
Living room	3.49%	0.62%	11.08%	1.71%

These statistics show that PBR-Rooms includes non-trivial metallic supervision across multiple indoor categories, with especially strong metallic coverage in kitchens. The metallic distribution is intentionally useful for evaluation and should not be interpreted as an estimate of real-world indoor metallic-material frequency.

Dataset Structure

The repository is organized into benchmark metadata, a review sample, and category-level corpus shards.

PBR-Rooms/
  README.md
  pbr_rooms_croissant.jsonld
  benchmark_manifest.csv

  split/
    scene_anonymous.csv

  review_sample/
    bathroom/
    bedroom/
    diningroom/
    kitchen/
    livingroom/

  bathroom/
    bathroom_part*.tar
  bedroom/
    bedroom_part*.tar
  diningroom/
    diningroom_part*.tar
  kitchen/
    kitchen_part*.tar
  livingroom/
    livingroom_part*.tar

The benchmark_manifest.csv file defines the official benchmark subset used for all reported experiments. Each row corresponds to one rendered RGB-to-X example and provides the relative paths to RGB and ground-truth channels.

The split/scene_anonymous.csv file contains the scene-level benchmark split without local paths or identifying information.

Category-level tar archives contain uploaded shards from the generated corpus. If not all category shards are present, they should be treated as a partial corpus release rather than the complete 39,000-image corpus.

Currently Uploaded Corpus Shards

At the time of the v0.1 review card, the repository may contain the following uploaded category-level shards:

Category	Uploaded shards	Release interpretation
Bathroom	bathroom_part1.tar–bathroom_part7.tar	Partial corpus shards
Bedroom	bedroom_part1.tar–bedroom_part7.tar	Partial corpus shards
Dining room	diningroom_part1.tar–diningroom_part7.tar	Partial corpus shards
Kitchen	kitchen_part1.tar–kitchen_part7.tar	Partial corpus shards
Living room	livingroom_part1.tar–livingroom_part7.tar	Partial corpus shards

If additional shards are uploaded, this table should be interpreted together with the repository file list and release notes. The official benchmark subset remains fixed by benchmark_manifest.csv and split/scene_anonymous.csv.

Expanded Scene Structure

After extracting a category-level archive, each expanded scene follows the structure below:

<scene_id>/
  V{view_id}_P{pose_id}_L{lighting_id}/
    Image/
    Depth/
    DepthNearWhite/
    SurfaceNormal/
    Albedo/
    Roughness/
    Metallic/
    ObjectSegmentation/

For a given scene_id, condition_id, and camera_id, all primary channels describe the same rendered view and are pixel-aligned at 1280 × 720 resolution. The current release uses camera_0.

The review sample is a lightweight subset for quick visual inspection. Depending on packaging, it may omit the anonymous scene-id directory level and place rendering-condition folders directly under each room category, for example:

review_sample/
  bathroom/
    V0_P0_L0/
      Image/
      Depth/
      DepthNearWhite/
      SurfaceNormal/
      Albedo/
      Roughness/
      Metallic/
      ObjectSegmentation/

Auxiliary folders such as Env/ and camview/, if present, are provided only for inspection/debugging and are not official benchmark targets. Generation marker files such as .render_done should be ignored.

Hugging Face Data Viewer Note

The Hugging Face Data Viewer may automatically interpret the review sample as an imagefolder-style dataset. This viewer is intended only for quick visual inspection and does not define the official benchmark schema. Official evaluation should use benchmark_manifest.csv, split/scene_anonymous.csv, and the released evaluation code.

Channel Specifications

Official quantitative evaluation should use numeric ground-truth files, preferably .exr or .npy when available. PNG files are intended mainly for visualization and quick inspection.

Channel	Folder	Official meaning	Official GT	Notes
RGB	Image	Rendered RGB input	.png / .exr	.png is suitable for RGB model input
Depth	Depth	Metric depth	.exr / .npy	Unit: meters
DepthNearWhite	DepthNearWhite	Near-white depth visualization	Visualization only	Brighter means closer
Surface normal	SurfaceNormal	Camera-space normal	.exr / .npy	Use released convention for evaluation
Albedo	Albedo	Intrinsic base color / diffuse reflectance	.exr	PNG previews, if present, are not official quantitative GT
Roughness	Roughness	PBR roughness	.exr / .npy	Range [0, 1]; low=smooth, high=rough
Metallic	Metallic	PBR metallic value	.exr / .npy	Range [0, 1]; 0=dielectric, 1=metallic
Segmentation	ObjectSegmentation	Object/semantic segmentation map	Numeric map when available	Auxiliary annotation; no public category label table in the current release

Depth ground truth is provided as metric depth in meters. The main benchmark uses affine-invariant relative-depth evaluation, but metric-depth evaluation can also be performed using the raw/source depth files.

Surface normals are provided in camera space. Predictions should be decoded or converted to match the released SurfaceNormal convention before angular-error evaluation.

The metallic channel represents a PBR material property, not a semantic object category. Objects of the same class may have different metallic values.

Benchmark Splits and Sampling

The full generated corpus contains 1,000 scenes and 39,000 images. The benchmark experiments use a reproducible scene-level sampling procedure rather than image-level random sampling.

The released benchmark split file is:

split/scene_anonymous.csv

It contains only non-identifying fields:

split,scene_type,scene

Benchmark Scene Sampling

Scenes are sampled independently within each room category:

• mainaxis: 4% of valid scenes per category, i.e., 8 scenes per category;
• stresstest: 5 additional non-overlapping scenes per category;
• sampling seed: 42.

Split	Bathroom	Bedroom	Dining room	Kitchen	Living room	Total scenes
mainaxis	8	8	8	8	8	40
stresstest	5	5	5	5	5	25
Total	13	13	13	13	13	65

The official benchmark evaluation subset contains 1,410 images:

Split	Rendering strategy	Images
mainaxis	40 scenes × 24 L0 camera views	960
stresstest	25 scenes × 3 cameras × 6 lighting conditions	450
Total	—	1,410

All compared models should use the same image list, preprocessing rules, evaluation-pixel selection, and metric implementation.

Benchmark Tasks

PBR-Rooms supports RGB-to-X benchmark tasks for evaluating geometry recovery, PBR material prediction, and multi-channel inverse rendering. Object segmentation is released as an auxiliary annotation, but it is not treated as a primary benchmark target in the current release.

Task	Input	Target	Main evaluation focus
RGB-to-depth	RGB	Depth	Scene geometry and near-far ordering
RGB-to-normal	RGB	SurfaceNormal	Local surface orientation
RGB-to-albedo	RGB	Albedo	Intrinsic material color
RGB-to-roughness	RGB	Roughness	Surface roughness prediction
RGB-to-metallic	RGB	Metallic	Metallic vs. non-metallic material prediction
Multi-channel inverse rendering	RGB	depth, normal, albedo, roughness, metallic	Joint geometry and material recovery

ObjectSegmentation may be used for qualitative analysis, data inspection, or auxiliary / prior-based ablation studies. Since the current release does not provide a public category label table, segmentation is not included as a primary quantitative benchmark task.

Detailed metric definitions are provided in the evaluation code and summarized below.

Evaluation Protocol

All predictions should be matched to the corresponding ground-truth files by scene, rendering condition, and camera. Predictions should be resized to 1280 × 720 if necessary. Evaluation is performed on finite, evaluable ground-truth pixels as implemented in the released evaluation code.

Channel	Metrics reported in the benchmark
Depth	AbsRel-AI, RMSE-AI, MAE-AI, δ1-AI, δ2-AI, Boundary F1
Surface normal	Mean angular error, Median angular error, Acc@11.25°, Acc@22.5°, Acc@30°
Albedo	MAE, PSNR, SSIM, LPIPS
Roughness	RMSE, MAE, SSIM, PSNR
Metallic	MAE, PSNR, SSIM, LPIPS
Segmentation	Auxiliary only in the current release; quantitative segmentation evaluation requires an explicit label mapping

For depth, the released ground truth is raw metric depth in meters. Model predictions may be metric depth or near-white relative-depth-style maps. The main benchmark evaluates depth under an affine-invariant single-image protocol by fitting an affine transformation from the prediction to the raw metric ground truth:

d_aligned = a * d_pred + b

where a and b are fitted on evaluable pixels. Metric-depth results, if reported, should be reported separately from affine-invariant relative-depth results.

For normal prediction, evaluation assumes the released camera-space SurfaceNormal convention. Predictions should be decoded or converted to this convention before angular-error evaluation.

For roughness and metallic prediction, outputs should be clamped or normalized to [0, 1] before evaluation if the model output range differs.

Evaluation Code

The anonymous evaluation code is provided in the OpenReview artifact. It is released separately under the MIT License and is used with benchmark_manifest.csv to reproduce the submitted benchmark metrics. The code implements prediction-to-ground-truth matching, resizing, affine-invariant depth alignment, normal angular-error evaluation, and image/material metrics.

Example command format:

python evaluate_depth.py \
  --manifest benchmark_manifest.csv \
  --gt_root ./PBR-Rooms \
  --pred_dir ./pred/depth \
  --out results_depth.json

After publication, a public repository link may be added here.

Loading and Usage

Download the dataset with the Hugging Face CLI:

huggingface-cli download pbr-rooms-benchmark/PBR-Rooms \
  --repo-type dataset \
  --local-dir ./PBR-Rooms

To download a single category:

huggingface-cli download pbr-rooms-benchmark/PBR-Rooms \
  --repo-type dataset \
  --include "kitchen/*" \
  --local-dir ./PBR-Rooms

Extract category archives with:

mkdir -p ./PBR-Rooms/kitchen
tar -xvf ./PBR-Rooms/kitchen/kitchen_part1.tar -C ./PBR-Rooms/kitchen/

Example loading code for the review sample:

from pathlib import Path
import imageio.v3 as iio
import numpy as np

root = Path("./PBR-Rooms/review_sample/diningroom")
condition = root / "V0_P0_L0"

rgb = iio.imread(next((condition / "Image" / "camera_0").glob("*.png")))
depth = iio.imread(next((condition / "Depth" / "camera_0").glob("*.exr"))).astype(np.float32)
metallic = iio.imread(next((condition / "Metallic" / "camera_0").glob("*.exr"))).astype(np.float32)

print(rgb.shape, depth.shape, metallic.shape)

For official evaluation, users should load files through benchmark_manifest.csv rather than manually enumerating folders.

Dataset Creation Pipeline

PBR-Rooms is generated using an Infinigen-based indoor scene generation and rendering pipeline. The pipeline procedurally creates indoor scenes, assigns physically based materials, renders RGB images, and exports aligned geometry, material, and segmentation maps.

The official ground-truth channels are exported directly from the rendering pipeline:

Channel	Official meaning
Depth	Metric depth in meters
SurfaceNormal	Camera-space surface normal
Albedo	Intrinsic base color / diffuse reflectance
Roughness	PBR roughness in [0, 1]
Metallic	PBR metallic value in [0, 1]
ObjectSegmentation	Object/semantic segmentation map

Scene validity filtering and channel completeness checks are applied during dataset construction. Invalid scenes are excluded from benchmark sampling.

Review Sample

Because the complete generated corpus is large, this repository includes a small review sample to allow inspection of RGB images, ground-truth channels, and file structure without downloading large tar shards.

review_sample/
  bathroom/
  bedroom/
  diningroom/
  kitchen/
  livingroom/

The review sample includes all primary channel folders. It is intended for quick visual inspection and should not be used as a replacement for the official benchmark subset defined by benchmark_manifest.csv.

Upstream Resources and Licensing

PBR-Rooms is generated using an Infinigen-based procedural indoor scene generation and rendering pipeline. The v0.1-review artifact does not intentionally redistribute real captured images, private indoor scans, personal data, or separate third-party public datasets. The released files are renderer-generated RGB images and renderer-exported ground-truth channels produced by the generation pipeline.

Upstream resource	Usage in PBR-Rooms	License / note
Infinigen / Infinigen-based indoor procedural generation pipeline	Procedural scene generation, material assignment, rendering, and ground-truth channel export	Infinigen is released under the BSD 3-Clause license. Users should also comply with the upstream Infinigen license when using the generation pipeline or derived tooling.

No additional public dataset such as 3D-FRONT, 3D-FUTURE, OpenRooms, Hypersim, or real-image indoor datasets is redistributed as part of this PBR-Rooms release. Baseline models or comparison datasets, if used in the paper, should be cited separately in the paper rather than treated as part of the PBR-Rooms dataset release.

A copy of this upstream-resource summary can also be placed at docs/upstream_licenses.md for easier reviewer inspection.

Responsible AI, Biases, and Limitations

PBR-Rooms is synthetic-only and does not intentionally include real people, real faces, real private homes, real addresses, biometric information, or personal user data.

The dataset may contain biases introduced by the procedural scene generator, object libraries, material assignment rules, rendering settings, and camera/lighting sampling strategy. It may not represent the full diversity of real indoor environments, architectural styles, geographic regions, cultural contexts, or household object distributions.

PBR-Rooms intentionally emphasizes reliable and non-trivial metallic supervision. This makes it useful for evaluating metallic prediction, but its metallic distribution should not be interpreted as an estimate of real-world indoor metallic-material frequency.

Results on PBR-Rooms should be interpreted as controlled synthetic benchmark results, not as proof of real-world deployment readiness. Additional evaluation on real captured images is required for claims about real-world generalization.

Out-of-scope uses include face recognition, person identification, surveillance, biometric inference, real-estate profiling, privacy-sensitive indoor reconstruction, autonomous driving, and outdoor scene understanding.

Croissant Metadata

This repository includes pbr_rooms_croissant.jsonld, which provides Croissant metadata for the dataset artifact. The file describes dataset-level metadata, released resources, split files, benchmark manifest, file formats, intended use, out-of-scope use, limitations, biases, synthetic data generation process, sensitive-information assessment, and release/maintenance plan.

The Croissant metadata includes both core fields and Responsible AI fields required for NeurIPS Evaluations & Datasets submissions.

Versioning and Maintenance

The anonymous review artifact is versioned as v0.1-review. It fixes the official benchmark subset, split file, and evaluation protocol used in the submitted paper.

Additional category-level corpus shards may be added as part of the versioned full release. Such additions will not change the official benchmark subset, benchmark manifest, split file, or reported evaluation protocol for v0.1-review.

Bug fixes to metadata, file paths, checksums, or documentation will be recorded through release notes. If any data file is corrected, the corresponding manifest and checksum files should be updated accordingly.

License

PBR-Rooms is released under the Creative Commons Attribution-NonCommercial 4.0 International license (CC BY-NC 4.0), unless otherwise specified.

The dataset is intended for academic research and non-commercial benchmarking. Users should provide appropriate attribution and cite the dataset paper when using PBR-Rooms.

The accompanying evaluation code is released separately under the MIT License.

The dataset is generated using an Infinigen-based procedural indoor scene generation and rendering pipeline. The upstream generation pipeline is Infinigen, which is released under the BSD 3-Clause license. Users should also follow the upstream Infinigen license when using or modifying the generation pipeline or derived tooling.

Citation

This dataset is released as an anonymous review artifact for a NeurIPS Evaluations & Datasets submission.

A formal citation will be added after publication.

Contact

For questions during anonymous review, please use the official review platform. After review, please use the Hugging Face discussion page or the public issue tracker if provided.