RoboClothA Large-Scale Real Cloth Material Dataset
for Neural Reflectance Reconstruction

Zhen (Colin) Li1*, Haoran Qin1,2*†, Francesco Di Sario3,1, Daniel Rebain4,5, Andrea Tagliasacchi1,5,6
1 Simon Fraser University    2 Westlake University    3 University of Torino    4 University of British Columbia    5 Wayve Labs    6 University of Toronto
* Co-first authors    † Work done during an internship at Simon Fraser University

arXiv Paper Code Dataset

Overview

Abstract

We present RoboCloth, a dataset of 500 real cloth materials captured under flexible, calibrated camera and lighting configurations, where both are robotically controlled during acquisition. Each material is imaged in approximately 580 high-dynamic range frames spanning a wide view-light hemisphere; camera poses are refined via structure-from-motion and aligned to the robot frame, yielding per-material sparse surface geometry and structured per-point reflectance observations ready for neural material training. We demonstrate the value of this data through a two-stage neural reconstruction pipeline: Stage 1 trains a shared MLP decoder on our corpus to learn cloth material priors; Stage 2 then freezes the decoder and fits a dense 2048×2048 latent texture to each target material. Evaluating on held-out materials from our own dataset and on the independent UBO2014 benchmark, we show that a decoder trained on our captures outperforms decoders trained on existing measured reflectance datasets and analytic PBR fitting, transferring to unseen acquisition systems without modification. The dataset, calibration files, observation tensors, and train/test splits are released under CC BY 4.0 on the Hugging Face Hub.

Pipeline

Stage 1 pipeline: sparse COLMAP points of several materials, per-point latent codes, normals and tangents, the shared transform T, and a jointly trained BRDF decoder
Stage 1 — learning a material prior. A single BRDF decoder, shared across all training materials, is optimized jointly (flame icon) with per-point tables of latent codes, normals, and tangents. N: normal, T: tangent, D: encoded local directions; each color denotes one material, with a few points drawn per material in every minibatch.
Stage 2 pipeline: a dense latent grid for one material, the parallax-aware lookup P, the shared transform T, and the frozen BRDF decoder
Stage 2 — dense material reconstruction. For a single target material, a dense latent-geometry texture is optimized (flame icon) through the parallax-aware lookup P, the shared transform 𝒯, and the frozen decoder (snowflake icon). Notation as in Stage 1.

Method

Texture the room with a trained material

Drag a material onto any sofa, pillow, cushion, the rug or the curtain. Tap a material, then any sofa, pillow, cushion, the rug or the curtain.

Loading the room…

Rendering note Each tile is a trained 2048×2048 latent texture, coloured by decoding every texel with the frozen decoder under a 45° light. Dropped on an object, it becomes the full material (relief, sheen and view dependence), which is path-traced on the piece you choose; every piece takes its own material. Each piece is pre-rendered in every material (3840 px, 1024 spp; 1920 px on phones and standard screens) and the combination is assembled live; the light a dressed piece throws onto the rest of the room is added as a smooth low-resolution correction.

Dataset Preview

The dataset is split into a train set and a test set following the official lists in globals/training_list_500.txt / globals/test_list_500.txt. 20 train and 5 test materials are shown as thumbnails below; every one of the remaining materials is linked further down so you can browse all 400 + 100 without leaving this page.

Train 20 of 400 shown

Loading preview manifest…

Test 5 of 100 shown

All material IDs — every train and test material, each linked to its viewer

Train all material IDs

Test all material IDs

Results

Compared with existing material datasets

Dataset Year Materials Cloth-specific Real-world Spatially varying Multiple views Flexible coverage
MERL main baseline 2003 100
UTIA 2014 150
UBO2014 main baseline 2014 84
UBOFAB19 main baseline 2019 378
Real-SVBRDF 2020 80
PhotoMat 2023 12,000
UMat 2023 2,000
OpenSVBRDF 2023 1,000
MatSynth 2024 4,069
RoboCloth ours 2026 500

Cloth-specific: the materials are fabrics. Real-world: measured from physical samples, not synthesised. Spatially varying: reflectance captured per surface point (a texture), not one BRDF per material. Multiple views: more than two distinct viewing directions. Flexible coverage: camera and light poses not confined to a fixed dome, gantry or scanner layout — RoboCloth is the only dataset with all five properties. Main baseline: UBO2014 is the test benchmark of the chart; MERL and UBOFAB19 train the competing decoders.

Reconstruction quality on UBO2014

A controlled protocol: the decoder architecture, renderer and Stage 2 procedure are held fixed across all comparisons, so differences in reconstruction quality are directly attributable to the training data — a decoder trained on RoboCloth, UBOFAB19 or MERL, fitted to 12 held-out UBO2014 materials; PBR is an analytic fit with no learned prior.

Horizontal bar chart. Mean PSNR on UBO2014 (12 held-out materials), best configuration per method. PSNR in dB, higher is better; the axis starts at 25 dB. RoboCloth 36.76, PBR 33.05, UBOFAB19 32.27, MERL 30.64.

Qualitative comparisons

Same protocol as above — the decoder, renderer and Stage 2 fitting are fixed and only the Stage 1 training data changes. Every recovered material is relit in a common scene: the same cloth mesh draped over a bar, with identical UV tiling, lighting and camera for every method, so any visible difference is due solely to the learned material. Click a figure to open it at full resolution.

UBO2014 — close-up on three held-out materials
UBOFAB19MERLPBRRoboClothGT
carpet12fabric02felt01
The MERL-trained decoder, trained only on spatially-uniform isotropic BRDFs, loses the weave structure and drifts in colour; the UBOFAB19-trained decoder recovers texture but shifts the overall hue (e.g. fabric02), and the analytic PBR baseline misses the fine sheen. Only the RoboCloth decoder matches the ground-truth colour while preserving both the specular highlight at the fold and the fuzzy weave detail. Each cell insets a magnified crop of the boxed region, taken from the full-resolution render; the number in each cell is the PSNR (dB) of the rendered image against the GT render, highest for RoboCloth in every cell.

Renderingpath traced at 1024×1024 (the carpets and fabric11 at 2048×2048), 512 spp, under an HDR environment map; the GT column is the measured UBO2014 BTF rendered in the same scene.

The full set — all 12 UBO2014 materials, the RoboCloth and UBOFAB19 test sets, and image-space comparisons against the captured photographs — is on its own page.

All qualitative results

Citation

@misc{li2026robocloth,
  title  = {RoboCloth: A Large-Scale Real Cloth Material Dataset for Neural Reflectance Reconstruction},
  author = {Li, Zhen and Qin, Haoran and Di Sario, Francesco and Rebain, Daniel and Tagliasacchi, Andrea},
  year   = {2026}
}