Welcome to the official repository for the Raformer project. This innovative model is engineered to address the challenge of removing wires from video sequences, showcasing its capabilities through extensive testing on our Wire Removal Video Dataset 2 (WRV2).

The WRV2 dataset is meticulously assembled to support developing and evaluating video inpainting algorithms aimed specifically at wire removal. This challenging task is critical for enhancing visual aesthetics in various scenes.

To download the WRV2 dataset, please visit this download link.

For those requiring higher resolution for detailed analysis, a 4K high-definition version of the original videos is available. Due to the large size of these files, approximately 2TB, it is not feasible to offer direct downloads. Please contact the authors directly to access these files or discuss potential delivery methods.

The dataset is organized as follows:

Wire Removal Video Datasets 2(WRV2)
|
├── train.json
├── test.json
├── JPEGImages
│   ├── yttl18-011_091
│   │   ├── 00000.jpg
│   │   ...
│   │   └── 0000N.jpg
│   ...
│   └── 12-042706_024
│       ├── 00000.jpg
│       ...
│       └── 0000N.jpg
│
└── test_masks
    ├── yttl18-011_091
    │   ├── 00000.png
    │   ...
    │   └── 0000N.png
    ...
    └── 12-042706_024
        ├── 00000.png
        ...
        └── 0000N.png

Annotations within this dataset are formatted as paletted binary images where the value 1 indicates the presence of a wire that needs to be inpainted, and 0 represents no wire. For a detailed look at the dataset's structure and contents, please refer to the example_for_WRV2.

The following image illustrates various scenes from the WRV2 dataset, highlighting the diversity and complexity of environments in our wire removal challenges:

Figure 1: Example Image 1 - Original Image from WRV2 Dataset

For a practical insight into the capabilities of the Raformer model and the challenges posed by wire artifacts in videos, we invite you to view our demonstration video. This video visually explains the preprocessing, challenges, and the effectiveness of the Raformer model in removing wires from video footage effectively.

If you are interested in exploring our previous dataset, the Wire Removal Video Dataset 1 (WRV1), please visit the following link for more information and resources:

If this research benefits your work or involves the use of this dataset, please consider citing the following paper:

@misc{ji2024raformer,
   title={Raformer: Redundancy-Aware Transformer for Video Wire Inpainting}, 
   author={Zhong Ji and Yimu Su and Yan Zhang and Jiacheng Hou and Yanwei Pang and Jungong Han},
   year={2024},
   eprint={2404.15802},
   archivePrefix={arXiv},
   primaryClass={cs.CV}
}

This project is licensed under the MIT License. Please refer to the LICENSE file for detailed terms.