README.md

---
license: apache-2.0
task_categories:
- image-to-image
tags:
- image-decomposition
- layered-image-editing
- object-removal
- image-matting
- image-inpainting
- computer-vision
- bounding-box
pretty_name: RevealLayer Open Dataset
size_categories:
- 100K<n<1M
language:
- en
---

# RevealLayer Open Dataset

RevealLayer Open is the open-source dataset accompanying **RevealLayer: Disentangling Hidden and Visible Layers via Occlusion-Aware Image Decomposition**.

Paper: https://arxiv.org/html/2605.11818v1 Accepted by ICML 2026

RevealLayer studies **box-guided layered image decomposition** for natural images. Given an RGB image and instance bounding boxes, the task is to decompose the scene into a clean background and object-level foreground layers, where each foreground layer is represented as RGBA.

This repository only redistributes data and annotations that are released by the RevealLayer authors. Third-party benchmark images and ground-truth annotations from AIM-500, RefMatte_RW100, and OBER-Test/ObjectClear are **not included** in this repository.

## License

The RevealLayer dataset, processed annotations, metadata, and scripts released in this repository are licensed under the **Apache License 2.0**.

Some evaluation metadata or conversion scripts may refer to third-party benchmarks, including AIM-500, RefMatte_RW100, and OBER-Test/ObjectClear. Their original images and ground-truth annotations are not redistributed here. Users should download those datasets from their official sources and follow the corresponding original licenses and usage terms.

## Repository Structure

A typical directory structure is:

```text
RevealLayer_open/
├── train/
│   ├── <sample_id>/
│   │   ├── full_image.png
│   │   ├── background.png
│   │   ├── layer_0.png
│   │   ├── layer_1.png
│   │   └── ...
│   └── metaData.json
│
├── Benchmark/
│   ├── RevealLayerBenchMark-200/
│   │   ├── <sample_id>/
│   │   │   ├── full_image.png
│   │   │   ├── background.png
│   │   │   ├── layer_0.png
│   │   │   └── ...
│   │   └── metaData.json
│   │
│   └── RevealLayerBenchMark-wild/
│       ├── <sample_id>/
│       │   └── full_image.png
│       └── metaData.json
│
└── README.md
```

The exact number of layers varies across samples.

## Metadata Format

Each split or benchmark subset contains a `metaData.json` file. It is a list of sample dictionaries.

### Training / Fully Annotated Samples

A fully annotated sample generally follows this format:

```json
{
  "imgid": "sample_id",
  "full_image": "sample_id/full_image.png",
  "background": "sample_id/background.png",
  "LayerInfoRaw": [
    "sample_id/layer_0.png",
    "sample_id/layer_1.png"
  ],
  "detections": [
    {
      "bbox": [x1, y1, x2, y2]
    }
  ]
}
```

Field meanings:

| Field | Type | Description |
| --- | --- | --- |
| `imgid` | string | Unique sample identifier. |
| `full_image` | string | Relative path to the original RGB image. |
| `background` | string | Relative path to the clean background image. |
| `LayerInfoRaw` | list[string] | Relative paths to object-level foreground RGBA layers. |
| `detections` | list[dict] | Instance bounding boxes used as box guidance. |
| `bbox` | list[number] | Bounding box in `[x1, y1, x2, y2]` format. |

The `detections` field only keeps bounding boxes. Labels and confidence scores are not required for the RevealLayer task and are not included.

### Wild Benchmark Samples

`RevealLayerBenchMark-wild` contains in-the-wild images with bounding-box annotations only. It does **not** include clean background ground truth or foreground RGBA ground truth.

A wild sample generally follows this format:

```json
{
  "imgid": "sample_id",
  "full_image": "sample_id/full_image.png",
  "background": "",
  "LayerInfoRaw": [],
  "detections": [
    {
      "bbox": [x1, y1, x2, y2]
    }
  ]
}
```

For `RevealLayerBenchMark-wild`, the `background` field may be an empty string and `LayerInfoRaw` may be empty. This indicates that no background or foreground-layer ground truth is provided.

## Benchmark Notes

### Included Benchmark Subsets

- **RevealLayerBenchMark-200**: a fully annotated benchmark subset for evaluating background reconstruction and foreground RGBA layer decomposition.
- **RevealLayerBenchMark-wild**: a wild-image benchmark subset with `full_image` and bounding boxes only. It is intended for qualitative and real-world robustness evaluation. It does not contain background or foreground-layer ground truth.

## Loading Example

```python
import json
from pathlib import Path
from PIL import Image

root = Path("RevealLayer_open/train")
metadata_path = root / "metaData.json"

with open(metadata_path, "r", encoding="utf-8") as f:
    samples = json.load(f)

sample = samples[0]
full_image = Image.open(root / sample["full_image"]).convert("RGB")

background = None
if sample.get("background"):
    background = Image.open(root / sample["background"]).convert("RGB")

layers = []
for layer_path in sample.get("LayerInfoRaw", []):
    layers.append(Image.open(root / layer_path).convert("RGBA"))

boxes = [det["bbox"] for det in sample.get("detections", [])]
```

## Data Usage Notes

- Paths in `metaData.json` are relative to the corresponding split or subset directory.
- Bounding boxes use `[x1, y1, x2, y2]` coordinates.
- Foreground layers are stored as RGBA images when ground truth is available.
- Some benchmark samples, especially wild images, may not contain background or foreground-layer ground truth.
- Third-party benchmark data are not redistributed in this repository. Users are responsible for complying with the original licenses when reproducing evaluations on those datasets.

## Citation

If you find this dataset useful, please cite:

```bibtex
@inproceedings{wang2026reveallayer,
  title={RevealLayer: Disentangling Hidden and Visible Layers via Occlusion-Aware Image Decomposition},
  author={Wang, Binhao and Zhao, Shihao and Cheng, Bo and Ji, Qiuyu and Ma, Yuhang and Wu, Liebucha and Liu, Shanyuan and Leng, Dawei and Yin, Yuhui},
  booktitle={International Conference on Machine Learning},
  year={2026}
}
```

#