README.md

---
pretty_name: PolarFree
tags:
- computer-vision
- reflection-removal
- polarization
- image-processing
- cvpr2025
license: cc-by-nc-4.0
datasets:
- polarfree
language:
- en
size_categories:
- 10K<n<100K
task_categories:
- image-to-image
---

# PolarFree: Polarization-based Reflection-Free Imaging

## Dataset Overview

PolarFree is a high-quality dataset designed for polarization-based reflection removal tasks, as introduced in the CVPR 2025 paper "PolarFree: Polarization-based Reflection-Free Imaging". The dataset aims to support tasks such as image reflection removal and image enhancement, particularly suitable for training and evaluating polarization-based image reflection removal models.

## Download Dataset
```
huggingface-cli download Mingde/PolaRGB --repo-type dataset --local-dir ./PolaRGB
```

## Dataset Structure

The dataset is organized as follows:

```
dataset/
├── train/
│   ├── easy (or hard)/                     # difficulty split
│   │   ├── input/
│   │   │   ├── 00/                         # scene 0
│   │   │   │   ├── 000_000.png
│   │   │   │   ├── 000_045.png
│   │   │   │   ├── 000_090.png
│   │   │   │   ├── 000_135.png
│   │   │   │   ├── 000_rgb.png
│   │   │   │   ├── 001_000.png
│   │   │   │   ├── 001_045.png
│   │   │   │   ├── 001_090.png
│   │   │   │   ├── 001_135.png
│   │   │   │   ├── 001_rgb.png
│   │   │   │   └── ...
│   │   │   ├── 01/                         # scene 1
│   │   │   │   ├── 000_000.png
│   │   │   │   ├── 000_045.png
│   │   │   │   ├── 000_090.png
│   │   │   │   ├── 000_135.png
│   │   │   │   ├── 000_rgb.png
│   │   │   │   ├── 001_000.png
│   │   │   │   ├── 001_045.png
│   │   │   │   ├── 001_090.png
│   │   │   │   ├── 001_135.png
│   │   │   │   ├── 001_rgb.png
│   │   │   │   └── ...
│   │   │   └── ...
│   │   ├── gt/
│   │   │   ├── 00/                         # scene 0
│   │   │   │   ├── 000_000.png
│   │   │   │   ├── 000_045.png
│   │   │   │   ├── 000_090.png
│   │   │   │   ├── 000_135.png
│   │   │   │   └── 000_rgb.png
│   │   │   ├── 01/                         # scene 1
│   │   │   │   ├── 000_000.png
│   │   │   │   ├── 000_045.png
│   │   │   │   ├── 000_090.png
│   │   │   │   ├── 000_135.png
│   │   │   │   └── 000_rgb.png
│   │   │   └── ...
│
├── test/
│   ├── input/
│   │   ├── 00/                             # scene 0
│   │   │   ├── 000_000.png
│   │   │   ├── 000_045.png
│   │   │   ├── 000_090.png
│   │   │   ├── 000_135.png
│   │   │   ├── 000_rgb.png
│   │   │   └── ...
│   │   ├── 01/                             # scene 1
│   │   │   ├── 000_000.png
│   │   │   ├── 000_045.png
│   │   │   ├── 000_090.png
│   │   │   ├── 000_135.png
│   │   │   ├── 000_rgb.png
│   │   │   └── ...
│   │   └── ...
│
│   ├── gt/
│   │   ├── 00/                             # scene 0
│   │   │   ├── 000_000.png
│   │   │   ├── 000_045.png
│   │   │   ├── 000_090.png
│   │   │   ├── 000_135.png
│   │   │   └── 000_rgb.png
│   │   ├── 01/                             # scene 1
│   │   │   ├── 000_000.png
│   │   │   ├── 000_045.png
│   │   │   ├── 000_090.png
│   │   │   ├── 000_135.png
│   │   │   └── 000_rgb.png
│   │   └── ...
└── ...
```

1. The dataset is divided into **train** and **test** subsets, and both follow a similar directory structure.
2. The **train** subset is further divided into **easy** and **hard** sets, which share the same internal structure. The **test** subset is not split by difficulty, but its structure is identical.
3. Taking the **test** set as an example, it contains two subfolders: **input** and **gt**. The *input* folder stores images with reflections, while the *gt* folder contains the corresponding clean images without reflections. Both folders include the same number of scenes (e.g., 00, 01, 02, ...).
4. In `test/input/00`, there are multiple images named in the format `xxx_yyy.png`. Here, **xxx** denotes the index of the captured sample within scene `00`. **yyy** may be one of {000, 045, 090, 135, rgb}: the first four represent polarization-based images, and `rgb` corresponds to the RGB image of the current scene.
5. In `test/gt/00`, there is a single image named `000_yyy.png`, where '000_rgb.png' serves as the ground truth for all captured samples in the corresponding `input/00` folder.

To train or evaluate a polarization-based reflection removal model, each sample pair can be constructed as follows. For a given scene (e.g., 00), the input consists of five images—000_000.png, 000_045.png, 000_090.png, 000_135.png, and 000_rgb.png—and the corresponding ground truth is gt/00/000_rgb.png. Similarly:

- input/00/{001_000.png, 001_045.png, 001_090.png, 001_135.png, 001_rgb.png} → gt/00/000_rgb.png

- input/00/{002_000.png, 002_045.png, 002_090.png, 002_135.png, 002_rgb.png} → gt/00/000_rgb.png

and so on for the remaining examples within the same scene.

If you want raw images, please find them at https://huggingface.co/datasets/Mingde/PolaRGB_raw or contact me via [email protected].

## Citation

If you use the PolarFree dataset in your research, please cite the following paper:

```bibtex
@inproceedings{polarfree2025,
  title={PolarFree: Polarization-based Reflection-Free Imaging},
  author={Yao, Mingde and Wang, Menglu and Tam, King-Man and Li, Lingen and Xue, Tianfan and Gu, Jinwei},
  booktitle={CVPR},
  year={2025}
}
```

For more information, please visit the project page:

- GitHub Repository: [https://github.com/mdyao/PolarFree](https://github.com/mdyao/PolarFree)
- Paper: [https://arxiv.org/abs/2503.18055](https://arxiv.org/abs/2503.18055)

---