This is a pytorch implementation for Progressively Complementary Network for Fisheye Image Rectification Using Appearance Flow.
This code contains two versions of the hyper-parameters. The first one is the implementation of node clustering task. The second one is the implementation of link prediction task.
- Linux or Windows
- Python 3
- Pytorch 1.5
For training the network, you need to download the perspective dataset Places2 or Coco. Then, move the downloaded images to
--data_prepare\picture
run
python data_prepare/get_dataset.py
to generate your fisheye dataset. The generated fisheye images and new GT will be placed in
--dataset\data\train
--dataset\gt\train
or
--dataset\data\test
--dataset\gt\test
Before training, make sure that the fisheye image has been placed in
--dataset/data/train
as well as corresponding GT is in
--dataset/gt/train
Update file paths in
--flist/dataset/train.flist
--flist/dataset/train_gt.flist
run
python train.py
If you want to use our pre-train model, you can download here.
Put the pre-train model in
--FISH-Net\release_model\pennet4_dataset_square256
placed test fisheye images in
--dataset/data/test
as well as corresponding GT is in (not necessary, but can be empty. You can placed the fisheye images to take up position.)
--dataset/gt/test
Update file paths in
--flist/dataset/test.flist
--flist/dataset/test_gt.flist
run
python test.py