tiptoez/MatE_polyhaven_extraction

---

license: cc0-1.0
pretty_name: Poly Haven Material Extraction Evaluation Set
size_categories:
- n<1K
task_categories:
- image-to-image
tags:
- computer-vision
- pbr materials
- textures
- polyhaven
- evaluation
- material-extraction

---

PolyHaven Material Extraction Evaluation Set

This dataset provides the PolyHaven evaluation set used in MatE: Material Extraction from Single-Image via Geometric Prior. It is designed for evaluating material extraction methods that recover physically based rendering (PBR) material maps from a single image and mask.

Dataset Contents

The dataset contains 216 samples. Each sample corresponds to one material and is stored in a separate directory named by its material ID, for example:

anti_skid_tiles/
asphalt_01/
brick_wall_09/

Each sample directory contains the following six PNG files:

image.png       # rendered input image for material extraction
mask.png        # binary foreground mask
albedo.png      # ground-truth diffuse/albedo map
normal_gl.png   # ground-truth normal map in OpenGL convention
roughness.png   # ground-truth roughness map
height.png      # ground-truth height/displacement map

All files are stored as 512 x 512 PNG images:

Evaluation Usage

For each sample, image.png and mask.png are intended to be used as the input to a material extraction method. The remaining maps, albedo.png, normal_gl.png, roughness.png, and height.png, serve as ground-truth targets for evaluation.

Optional non-core maps such as ambient occlusion, metallic, specular, and material masks are not included. They were removed so that every evaluation sample follows a consistent file structure and contains the same set of target maps.

Paper

This evaluation set is associated with:

MatE: Material Extraction from Single-Image via Geometric Prior Zeyu Zhang, Wei Zhai, Jian Yang, Yang Cao arXiv: 2512.18312

If this evaluation set is useful for your work, please consider citing our paper:

@article{zhang2025mate,
  title={MatE: Material Extraction from Single-Image via Geometric Prior},
  author={Zhang, Zeyu and Zhai, Wei and Yang, Jian and Cao, Yang},
  journal={arXiv preprint arXiv:2512.18312},
  year={2025}
}

License

This dataset is released under the CC0-1.0 license.