---
library_name: pytorch
pipeline_tag: image-to-image
tags:
  - polarization
  - demosaicking
  - shape-from-polarization
  - reflection-removal
---

# PolarAPP checkpoints

This repository contains the released checkpoints for the two downstream tasks in [PolarAPP](https://arxiv.org/abs/2603.23071):

- `SfP/`: full-resolution shape from polarization with a TaskNet designed specifically for PolarAPP.
- `DfP/`: full-resolution de-reflection from polarization with a PolarFree-based TaskNet.

## Files

```text
SfP/
|-- DemNet/DemNet.pth
|-- TaskNet/TaskNet.pth
`-- FANet/FANet.pth

DfP/
|-- DemNet/DemNet.pth
|-- TaskNet/TaskNet.pth
`-- FANet/FANet.pth
```

`DemNet` reconstructs full-resolution polarization observations. `TaskNet` performs the downstream task. `FANet` supports feature alignment during training.

## Download

```python
from huggingface_hub import snapshot_download

checkpoint_root = snapshot_download("Roydon728/PolarAPP")
```

Pass the task subdirectory to the corresponding inference command:

```bash
cd SfP
python infer.py --input-dir ./Datasets/Testsets --ckpt-dir <snapshot>/SfP

cd DfP
python infer.py --input-dir ./Datasets/PolaRGB --ckpt-dir <snapshot>/DfP \
  --polarfree-checkpoint-dir ./experiments/checkpoints/polarfree
```

The DfP workflow loads the PolarFree diffusion-prior files separately using their upstream filenames.

## Data

- SfP dataset links are distributed through [SfPUEL](https://github.com/YouweiLyu/SfPUEL). Follow the dataset terms and citation guidance from the provider.
- DfP uses [PolaRGB](https://huggingface.co/datasets/Mingde/PolaRGB), distributed under CC BY-NC 4.0.

## Integrity

SHA-256 checksums are provided in `SHA256SUMS.txt`.

## Citation

```bibtex
@article{luo2026polarapp,
  title   = {PolarAPP: Beyond Polarization Demosaicking for Polarimetric Applications},
  author  = {Luo, Yidong and Li, Chenggong and Song, Yunfeng and Wang, Ping and Shi, Boxin and Zhang, Junchao and Yuan, Xin},
  journal = {arXiv preprint arXiv:2603.23071},
  year    = {2026}
}
```