README.md

---
license: mit
pretty_name: MatPredict
task_categories:
- image-to-image
- image-segmentation
tags:
- synthetic-data
- inverse-rendering
- material-segmentation
- material-properties
- pbr
- robotics
- indoor-scenes
---

# MatPredict

MatPredict is a synthetic dataset for material-centric scene understanding. It supports two main tasks:

1. **Inverse rendering**: predict material properties such as albedo, roughness, and metallic maps from RGB images.
2. **Material segmentation**: predict material regions or material classes from RGB images.

The dataset contains rendered object variants with paired RGB images, material property maps, segmentation labels, camera transforms, and metadata.

## Dataset Structure

```text
MatPredict/
  material_segmentation_map.yaml
  config/
    object_disjoint_v1.yaml
    variance_disjoint_v1.yaml

  <object_name>/
    <variant_name>/
      images/        # RGB input images
      albedo/        # base color targets
      ORM/           # packed material map; roughness=G, metallic=B
      label/         # material segmentation labels
      depth/
      normal_mat/
      normal_obj/
      transforms.json
      metadata.json
      material_segmentation_map.json
```

## Tasks

### Inverse Rendering

Input:

```text
images/*.png
```

Targets:

```text
albedo/*.png
ORM/*.png
```

### Material Segmentation

Input:

```text
images/*.png
```

Target:

```text
label/*.png
```

## Splits

The dataset includes two split files:

- `config/object_disjoint_v1.yaml`: train, validation, and test sets use disjoint object identities.
- `config/variance_disjoint_v1.yaml`: train, validation, and test sets use disjoint material/rendering variants.

Both split files store relative sample ids in the form:

```text
<object_name>/<variant_name>/<frame_id>
```