Omnidirectional Image Super-resolution via Bi-projection Fusion (AAAI 2024) [PDF]

Jiangang Wang | Yuning Cui | Yawen Li | Wenqi Ren | Xiaochun Cao

Shenzhen Campus of Sun Yat-sen University
Technical University of Munich
Beijing University of Posts and Telecommunications

⭐ If BPOSR is helpful to your images or projects, please consider starring this repo. Thank you! 🤗

🚩 Accepted by AAAI 2024

🔎 Poster

---

⚙️ Dependencies and Installation

To get started, follow these steps to clone the repository, set up your environment, and install the required dependencies:

## Clone this repository
git clone https://github.com/W-JG/BPOSR
cd BPOSR

## Create a conda environment
conda create -n BPOSR python=3.10
conda activate BPOSR

## Install dependencies
pip install -r requirements.txt
python setup.py develop

---

🗂️ Prepare Training and Testing Data

Step 1: Download LAU Dataset

Download the LAU dataset from this link and place it in the Dataset directory.
To extract the data, run:

cd Dataset
unrar x lau_dataset.rar

Step 2: Clean the Data

To clean the dataset, similar to OSRT, run the following command:

python Data_prepare/make_clean_lau_dataset.py

This will create a lau_dataset_clean folder inside the Dataset directory.

---

🚀 Quick Inference

Step 1: Download Pretrained Models

Download the pretrained weights from this link and place them in the BPOSR/experiments folder.

Step 2: Run the Testing Command

Use the following commands to test with different scales:

CUDA_VISIBLE_DEVICES=0 python basicsr/test.py -opt ./options/BPOSR_test/test_BPOSR_x4.yml
CUDA_VISIBLE_DEVICES=1 python basicsr/test.py -opt ./options/BPOSR_test/test_BPOSR_x8.yml
CUDA_VISIBLE_DEVICES=2 python basicsr/test.py -opt ./options/BPOSR_test/test_BPOSR_x16.yml

---

🌈 Training

Single GPU

For single-GPU training, run the following command:

PYTHONPATH="./:${PYTHONPATH}" CUDA_VISIBLE_DEVICES=0 python -W ignore basicsr/train.py -opt options/BPOSR_train/train_BPOSR_SW_SRx4.yml

Multi-GPU

For multi-GPU training, use the following command:

PYTHONPATH="./:${PYTHONPATH}" CUDA_VISIBLE_DEVICES=0,1,2,3 python -W ignore -m torch.distributed.launch --nproc_per_node=4 --master_port=4321 --use_env basicsr/train.py -opt options/BPOSR_train/train_BPOSR_SW_SRx4.yml --launcher pytorch

Note: Be sure to modify the train_BPOSR_SW_SRx4.yml file and set num_gpu: 4 accordingly.

---

📷 Presentation

For a quick understanding of the project, you can check out the poster and presentation slides:

• English
• Chinese

---

❤️ Acknowledgments

This project is built upon BasicSR.
The core code for BPOSR is located in the following files:

• basicsr/archs/BPOSR_arch.py
• basicsr/metrics/odi_metric.py
• basicsr/data/OmniImage_dataset.py
• basicsr/data/Omni_transforms.py

---

📧 Contact

For any questions, feel free to reach out to:
Email: [email protected]

---

🎓 Citations

If this code is helpful for your research or projects, please consider citing our paper. Here's the BibTeX reference:

@inproceedings{wang2024omnidirectional,
  title={Omnidirectional Image Super-resolution via Bi-projection Fusion},
  author={Wang, Jiangang and Cui, Yuning and Li, Yawen and Ren, Wenqi and Cao, Xiaochun},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={38},
  number={6},
  pages={5454--5462},
  year={2024}
}

---

🎫 License

This project is released under the MIT License.

---

Apology

We apologize for the delayed release of the code. After the paper was submitted, the authors changed affiliations, which made it challenging to access the original data. We appreciate your understanding.

---

Statistics