Dataset Viewer
Auto-converted to Parquet Duplicate
sample_id
stringlengths
36
36
depth
array 2D
occlusion
array 2D
ltm_rows
listlengths
200k
1.43M
ltm_cols
listlengths
200k
1.43M
ltm_values
listlengths
200k
1.43M
141bd91a-06c7-4f63-a9f7-f7f4e5539320
[[0.999000072479248,0.999000072479248,0.9989999532699585,0.9990000128746033,0.998999834060669,0.9990(...TRUNCATED)
[[true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,tru(...TRUNCATED)
[0,0,0,0,1,1,2,2,2,3,3,3,4,4,4,4,5,5,6,6,6,7,7,7,7,7,8,8,8,8,8,8,8,9,9,9,9,10,10,10,10,11,11,11,11,1(...TRUNCATED)
[64514,64515,64770,64771,64515,64771,61976,64516,64772,64517,64772,64773,64517,64518,64773,64774,645(...TRUNCATED)
[3894,177,1416,177,3894,1770,32,3363,2301,3894,177,1593,177,3009,177,2301,3717,1947,354,2301,3009,32(...TRUNCATED)
0f45ba08-0a40-4c28-9526-72029ee3ac8d
[[0.9990001320838928,0.9990001916885376,0.9989999532699585,0.9990000128746033,0.998999834060669,0.99(...TRUNCATED)
[[true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,tru(...TRUNCATED)
[0,0,0,0,1,1,2,2,2,3,3,3,4,4,4,4,5,5,6,6,6,7,7,7,7,8,8,8,8,8,9,9,9,9,10,10,10,10,11,11,11,11,12,12,1(...TRUNCATED)
[64514,64515,64770,64771,64515,64771,35899,64516,64772,64517,64772,64773,64517,64518,64773,64774,645(...TRUNCATED)
[3894,177,1416,177,3894,1770,63,3363,2301,3894,177,1593,177,3009,177,2301,3717,1947,354,2301,3009,12(...TRUNCATED)
f280814d-c2a5-4320-a13f-1669798f7938
[[0.999000072479248,0.9990000128746033,0.9989999532699585,0.9990000128746033,0.998999834060669,0.999(...TRUNCATED)
[[true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,tru(...TRUNCATED)
[0,0,0,0,1,1,1,2,2,3,3,3,4,4,4,4,5,5,6,6,6,6,7,7,7,7,8,8,8,8,9,9,9,9,10,10,10,10,10,11,11,11,11,12,1(...TRUNCATED)
[64514,64515,64770,64771,56879,64515,64771,64516,64772,64517,64772,64773,64517,64518,64773,64774,645(...TRUNCATED)
[3894,177,1416,177,32,3894,1770,3363,2301,3894,177,1593,177,3009,177,2301,3717,1947,32,354,2301,3009(...TRUNCATED)
40dae031-73fa-405c-83cb-f9bc1b320367
[[0.9989999532699585,0.9990000128746033,0.9989999532699585,0.9990000128746033,0.998999834060669,0.99(...TRUNCATED)
[[true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,tru(...TRUNCATED)
[0,0,0,0,1,1,2,2,3,3,3,4,4,4,4,5,5,6,6,6,6,7,7,7,7,8,8,8,8,8,9,9,9,9,10,10,10,10,10,10,11,11,11,11,1(...TRUNCATED)
[64514,64515,64770,64771,64515,64771,64516,64772,64517,64772,64773,64517,64518,64773,64774,64519,647(...TRUNCATED)
[3894,177,1416,177,3894,1770,3363,2301,3894,177,1593,177,3009,177,2301,3717,1947,33,354,2301,3009,12(...TRUNCATED)
7a17dcab-b85f-4906-a9da-3a9bf32ce356
[[0.999000072479248,0.999000072479248,0.9989999532699585,0.9990000128746033,0.998999834060669,0.9990(...TRUNCATED)
[[true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,tru(...TRUNCATED)
[0,0,0,0,1,1,2,2,3,3,3,4,4,4,4,5,5,6,6,6,7,7,7,7,8,8,8,8,8,9,9,9,9,9,10,10,10,10,10,11,11,11,11,12,1(...TRUNCATED)
[64514,64515,64770,64771,64515,64771,64516,64772,64517,64772,64773,64517,64518,64773,64774,64519,647(...TRUNCATED)
[3894,177,1416,177,3894,1770,3363,2301,3894,177,1593,177,3009,177,2301,3717,1947,354,2301,3009,1239,(...TRUNCATED)
1c3986df-cc9a-473c-bfa8-b729f84f0e62
[[0.999000072479248,0.9990000128746033,0.9989999532699585,0.9990000128746033,0.998999834060669,0.999(...TRUNCATED)
[[true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,tru(...TRUNCATED)
[0,0,0,0,0,1,1,2,2,2,2,3,3,3,4,4,4,4,5,5,6,6,6,7,7,7,7,7,7,8,8,8,8,8,9,9,9,9,9,10,10,10,10,11,11,11,(...TRUNCATED)
[61301,64514,64515,64770,64771,64515,64771,49016,56106,64516,64772,64517,64772,64773,64517,64518,647(...TRUNCATED)
[33,3894,177,1416,177,3894,1770,33,32,3363,2301,3894,177,1593,177,3009,177,2301,3717,1947,354,2301,3(...TRUNCATED)
14a43bf8-e668-437c-abfe-2cc46efbd503
[[0.9989999532699585,0.9989998936653137,0.9990000128746033,0.9989999532699585,0.998999834060669,0.99(...TRUNCATED)
[[true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,tru(...TRUNCATED)
[0,0,0,0,0,0,1,1,2,2,2,3,3,3,4,4,4,4,4,4,5,5,6,6,6,6,7,7,7,7,7,8,8,8,8,8,8,8,8,8,9,9,9,9,9,10,10,10,(...TRUNCATED)
[64514,64515,64620,64770,64771,65029,64515,64771,64011,64516,64772,64517,64772,64773,64517,64518,647(...TRUNCATED)
[3894,177,33,1416,177,32,3894,1770,32,3363,2301,3894,177,1593,177,3009,177,2301,32,32,3717,1947,32,3(...TRUNCATED)
ec39d3f3-0eaf-4da9-aa69-242f72c531e3
[[0.999000072479248,0.999000072479248,0.9989999532699585,0.9989999532699585,0.998999834060669,0.9990(...TRUNCATED)
[[true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,tru(...TRUNCATED)
[0,0,0,0,1,1,1,2,2,3,3,3,4,4,4,4,5,5,6,6,6,7,7,7,7,8,8,8,8,8,9,9,9,9,9,10,10,10,10,11,11,11,11,12,12(...TRUNCATED)
[64514,64515,64770,64771,53060,64515,64771,64516,64772,64517,64772,64773,64517,64518,64773,64774,645(...TRUNCATED)
[3894,177,1416,177,32,3894,1770,3363,2301,3894,177,1593,177,3009,177,2301,3717,1947,354,2301,3009,12(...TRUNCATED)
1c4a3558-f970-4ff9-ac0b-4123d31eb459
[[0.999000072479248,0.999000072479248,0.9989999532699585,0.9990000128746033,0.998999834060669,0.9990(...TRUNCATED)
[[true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,tru(...TRUNCATED)
[0,0,0,0,1,1,2,2,3,3,3,3,4,4,4,4,5,5,6,6,6,7,7,7,7,8,8,8,8,9,9,9,9,10,10,10,10,10,11,11,11,11,11,12,(...TRUNCATED)
[64514,64515,64770,64771,64515,64771,64516,64772,55553,64517,64772,64773,64517,64518,64773,64774,645(...TRUNCATED)
[3894,177,1416,177,3894,1770,3363,2301,32,3894,177,1593,177,3009,177,2301,3717,1947,354,2301,3009,12(...TRUNCATED)
23785fd3-a04d-4046-989c-d32e7ff5921d
[[0.999000072479248,0.999000072479248,0.9989998936653137,0.9989999532699585,0.998999834060669,0.9990(...TRUNCATED)
[[true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,tru(...TRUNCATED)
[0,0,0,0,1,1,2,2,2,3,3,3,4,4,4,4,5,5,6,6,6,7,7,7,7,7,8,8,8,8,8,8,8,9,9,9,9,10,10,10,10,10,11,11,11,1(...TRUNCATED)
[64514,64515,64770,64771,64515,64771,29021,64516,64772,64517,64772,64773,64517,64518,64773,64774,645(...TRUNCATED)
[3894,177,1416,177,3894,1770,12,3363,2301,3894,177,1593,177,3009,177,2301,3717,1947,354,2301,3009,32(...TRUNCATED)
End of preview. Expand in Data Studio

ABC Light Transport 256

A synthetic dataset for structured-light 3D scanning made with a projector–camera pair. Every sample is a CAD object from the ABC dataset placed on a table and rendered with physically based path tracing in a modified Mitsuba 3. Each sample stores:

  • the full light transport matrix (LTM) between the 256×256 projector and the 256×256 camera, including global illumination (interreflections). It lets you render the camera image for any projected pattern with a single sparse matrix–vector product.
  • the ground-truth depth map seen by the camera.
  • a projector-shadow mask that marks which camera pixels are directly lit by the projector.

Because the LTM is stored instead of images for a fixed pattern set, the dataset can be used to train and evaluate arbitrary (including learned) structured-light patterns and decoders.

There are two configurations. They use the same 100,000 meshes and differ only in the object's material:

config object material train test size
diffuse Lambertian, reflectance 0.6 90,000 10,000 ~94 GB
metalic Rough conductor, aluminium (Al), GGX roughness α = 0.15 90,000 10,000 ~108 GB

Loading

from datasets import load_dataset

ds = load_dataset("leon3428/abc-light-transport-256", "diffuse", split="train", streaming=True)
ds = ds.with_format("numpy")
sample = next(iter(ds))

Simulating a camera image for a projector pattern

import numpy as np
import scipy.sparse as sp

N = 256 * 256
ltm = sp.csr_matrix(
    (sample["ltm_values"].astype(np.float32) / 65535.0,
     (sample["ltm_rows"], sample["ltm_cols"])),
    shape=(N, N),  # (camera pixels, projector pixels)
)

pattern = np.random.rand(256, 256).astype(np.float32)  # projector image in [0, 1]
camera = (ltm @ pattern.reshape(-1)).reshape(256, 256)  # simulated camera image

Fields

field type description
sample_id string Random UUID of the sample (folder name in the original release).
depth float32 [256, 256] Camera-space z-depth of the first surface hit (not Euclidean ray length), in scene units. The camera sits 1 unit from the table, so values are ≈ 0.83–1.0. Averaged over the pixel's samples, so depth at silhouette edges is blended.
occlusion bool [256, 256] True where the camera pixel is directly lit by the projector (direct-only render > 1e-3); False in projector shadow.
ltm_rows uint32 [nnz] LTM row = camera pixel index y * 256 + x (row-major). Sorted ascending.
ltm_cols uint32 [nnz] LTM column = projector pixel index v * 256 + u (row-major); sorted within each row.
ltm_values uint16 [nnz] Fixed-point transport coefficient; divide by 65535 for the float value.

The LTM is a 65536 × 65536 sparse matrix in COO format with no duplicate (row, col) pairs. A typical sample has a few hundred thousand nonzeros (about 0.24M–1.2M for diffuse and 0.23M–3.2M for metalic).

Generation details

Scenes were built and rendered with dataset_generator from the structured-light project, using Mitsuba 3's cuda_mono variant, which renders single-channel (monochrome) light.

Scene

  • Camera: perspective, 256×256, horizontal FOV 10°, at (0, 0, -1), looking along +z.
  • Projector: Mitsuba projector emitter, 256×256, horizontal FOV 10°, irradiance 1.0, scale 0.3. It sits at (-0.1, 0, -1) (a 10 cm horizontal baseline) and is rotated 5.7° about the y-axis toward the camera's view.
  • Table: a fixed background plane (background.ply) with a diffuse material, reflectance 0.3.
  • Object: one ABC mesh per sample, with a two-sided material (see the configs above). Each sample gets a random pose:
    • uniform scale so the largest bounding-box side is 0.12–0.17 units
    • random x/y offset in ±0.05
    • random rotation of 0–360° about the camera axis
    • placed so that it rests on the table

Light transport matrix

  • Each camera pixel is path traced at 32 samples per pixel with a custom light-transport integrator (up to 4 path vertices, emitter sampling with MIS). Every path that reaches the projector adds its contribution to the entry (camera pixel, projector pixel it lands in).
  • A single path-sample contribution (already divided by the spp count) is discarded if it is ≤ 1e-4. Otherwise it is clamped to 1, quantized to round(value * 65535) as uint16, and then summed with the other samples that hit the same (row, col) entry.

Depth comes from a zdepth integrator at 32 spp. Occlusion comes from a path tracer limited to direct illumination (max_depth = 2), thresholded at 1e-3.

Meshes: the first 100,000 of 105,903 meshes taken from the ABC dataset (OBJ release), each recentred on its bounding-box centre. Each mesh appears once per config, with an independently sampled pose.

Split: 90% train and 10% test, made separately for each config from the directory listing order. The split is per sample, not per mesh across configs, so a mesh in diffuse/test may appear in metalic/train.

Conversion from the original format

The original release stored each sample as a folder <sample_id>/{depth,ltm,occlusion}.h5. The conversion to Parquet is lossless and keeps the train/test assignment and order from split.json:

  • depth.h5:/depth and occlusion.h5:/occlusion (256×256×1) → depth, occlusion, with the trailing size-1 channel dropped
  • ltm.h5:/rows, /cols, /values → ltm_rows, ltm_cols, ltm_values
  • the ltm.h5 attributes camera_{height,width} and projector_{height,width} were 256 for every sample, so they are not stored per row

The original split.json and chunks.json are kept in metadata/<config>/. chunks.json is generation bookkeeping that lists the internal IDs of the source meshes. Sample IDs are random UUIDs, and no sample-to-mesh mapping was recorded. The conversion and verification scripts are in scripts/.

License

The rendered data in this repository (light transport matrices, depth maps, occlusion masks) is released under CC BY 4.0.

The objects in the scenes are CAD models from the ABC dataset, and their license is separate. Copyright in each model belongs to its creator, and the models are licensed under the Onshape Terms of Use, section 1.g.ii:

  • Most models are covered by a worldwide, royalty-free, non-exclusive grant to "use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies".
  • Public documents created before August 7, 2018 with a LICENSE tab, and those owned by non-Free-Plan users with such a tab, may reserve greater rights.

The ABC dataset does not record per-model license tabs. If your use depends on the license of a specific mesh, check the ABC/Onshape terms. Neither this dataset nor ABC comes with any warranty.

Citation

If you use this dataset, please also cite the ABC dataset and Mitsuba 3:

@inproceedings{Koch_2019_CVPR,
  title     = {ABC: A Big CAD Model Dataset For Geometric Deep Learning},
  author    = {Koch, Sebastian and Matveev, Albert and Jiang, Zhongshi and Williams, Francis and Artemov, Alexey and Burnaev, Evgeny and Alexa, Marc and Zorin, Denis and Panozzo, Daniele},
  booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2019}
}

@software{Mitsuba3,
  title   = {Mitsuba 3 renderer},
  author  = {Wenzel Jakob and Sébastien Speierer and Nicolas Roussel and Merlin Nimier-David and Delio Vicini and Tizian Zeltner and Baptiste Nicolet and Miguel Crespo and Vincent Leroy and Ziyi Zhang},
  note    = {https://mitsuba-renderer.org},
  version = {3.1.1},
  year    = {2022}
}
Downloads last month
96