koalapenguin/cloth-brdf

Cloth BRDF Dataset (Multi-View Multi-Light HDR)

Multi-view multi-light HDR captures of cloth materials with associated 3D geometry, designed for inverse rendering and BRDF estimation.

Snapshot: 500 materials, 3.62 TB total (3,622 GB), 9,853 files.

• License: CC-BY-4.0 (see LICENSE)
• Croissant metadata: croissant.jsonld
• Sample subset (≤4 GB): sample/ — for quick reviewer inspection
• Globals: globals/ — calibration, sample-size table, train/test splits
• Examples: examples/load_material.py

Dataset summary

Cloth-BRDF is a large-scale dataset of densely sampled cloth material appearances captured under controlled multi-view, multi-light conditions. Each material is represented by a structured set of HDR observations, per-camera and per-light pose information, and a sparse 3D point cloud reconstructed from the calibration imagery.

The dataset is intended to enable BRDF / SVBRDF estimation, photometric stereo benchmarks, and inverse-rendering research that requires reliable multi-view, multi-light ground truth on real-world cloth samples.

Repository layout

materials/{id}/
  hdr.tar                            # 500-585 16-bit HDR PNGs (multi-view, multi-light)
  observations_structured.npz        # xyz, point_ids, rgbs, cam_pos, light_pos arrays
  point_positions.npz                # sparse 3D point cloud
  rotated_camera.json                # per-view camera poses (registered to rig frame)
  scan_log.json                      # per-scan camera + light pose log
  point_metadata.json                # observation count, point count
  bbox.json                          # sample bounding box
  hdr_crop_bboxes.json               # per-image crop polygon + dilation metadata
  unmatched_scan_ids.json            # scans without camera/light correspondence

globals/
  camera_factor.json                 # per-camera intensity correction factors
  emitter_calibration.json           # light emitter intensity profile (degrees)
  sample_size.json                   # material physical-size groups
  training_list_{N}.txt            # training splits at N = 100, 300, 442, 500
  test_list_{N}.txt                # corresponding test splits

sample/                              # ≤4 GB downsampled subset for reviewer inspection
examples/load_material.py            # minimal loader

croissant.jsonld                     # MLCommons Croissant metadata (core + RAI)
LICENSE                              # CC-BY-4.0

Data collection

Captures performed with a custom rig combining a robot-arm-mounted camera, an array of LED light sources at calibrated positions, and a sample-holding platform with markers for pose recovery. Each material sample is mounted flat and imaged from <FILL_IN: number> camera viewpoints under <FILL_IN: number> lighting conditions, producing roughly 500-585 16-bit HDR PNG images per material. A sparse 3D reconstruction (COLMAP) recovers point geometry and registers cameras into the rig coordinate frame. Per-pixel observations are then assembled into a structured npz with (point, camera, light) indexing.

Annotations

No human annotation. All metadata (camera poses, light positions, per-pixel observations, point cloud) is derived from the calibration pipeline.

Loading

pip install huggingface_hub numpy pillow
python examples/load_material.py --mid 0

For batched / streamed loading, treat each materials/{id}/hdr.tar as a WebDataset shard.

Splits

The training/test split scales with the number of materials desired. Choose the variant matching your experiment:

Split file	Materials	Use case
globals/training_list_100.txt + test_list_100.txt	100	small-scale ablations
globals/training_list_300.txt + test_list_300.txt	300	medium-scale benchmarks
globals/training_list_442.txt + test_list_442.txt	442	full prior to material 400 (small-npz outlier)
globals/training_list_500.txt + test_list_500.txt	500	full dataset

Limitations

• Single capture rig: rig-specific calibration assumptions (lens model, light intensity profile, geometric layout) are baked into the data.
• Cloth deformations are not modelled — samples are flat-mounted and imaged in a planar configuration.
• Specular highlights at grazing angles may be clipped despite the 16-bit HDR encoding.
• Sparse 3D points (typically <FILL_IN: range> per material) are derived from feature-matched calibration imagery rather than dense scanning.

Biases

• Material distribution is biased toward fabrics readily available in <FILL_IN: e.g. North American retail / lab partner suppliers>; not a representative cross-section of global textile diversity.
• Lighting hemisphere only (no transmissive setups, no sub-surface scattering captures).
• HDR capture, while wide-range, may saturate on very specular or very dark materials.

Personal / sensitive information

None. Data consists exclusively of cloth material captures. No people, no faces, no identifiable subjects, no personally-identifying metadata.

Intended use cases

• BRDF / SVBRDF estimation
• Photometric stereo benchmarks
• Multi-view inverse rendering
• Neural appearance models conditioned on geometry + lighting
• Material classification or retrieval research

Social impact

• Intended for graphics, vision, and inverse-rendering research.
• No known dual-use risk: cloth material captures are physical-object measurements with no human subjects, no personally-identifying metadata, and no operational security implications.
• Indirect downstream uses might include realistic cloth rendering for games, films, or virtual try-on; the dataset itself does not enable surveillance or harm.

Citation

@misc{clothbrdf2026,
  title  = {<FILL_IN: paper title>},
  author = {<FILL_IN: anonymous during review>},
  year   = {2026},
  note   = {NeurIPS 2026 Datasets \& Benchmarks Track submission},
  url    = {https://huggingface.co/datasets/koalapenguin/cloth-brdf}
}

Provenance

Pipeline scripts are under scripts/dataset_submission/ in the source repository:

• Capture: capture/capture_pipeline_fixed_center.py
• Reconstruction: COLMAP feature matching + triangulation
• HDR cropping: scripts/dataset_submission/crop_hdr_by_mask.py
• Observation structuring: scripts/dataset_submission/process_all_materials.py
• Upload: scripts/dataset_submission/upload_hf_debug.py

Detailed Croissant prov:wasGeneratedBy records are in croissant.jsonld.