Global Convolutional Neural Networks With Self-Attention for Fisheye Image Rectification (IEEE Access)

Introduction

This is a TensorFlow implementation for Global Convolutional Neural Networks With Self-Attention for Fisheye Image Rectification.
This work has been published in IEEE Access.

Architecture

Installation

1. Clone the project

   git clone https://github.com/byunghyun23/GSAFE

2. Install the prerequisites

   pip install -r requirements.txt

Dataset

Before training a model, you need to download the dataset here or full Places2.
Then, move the downloaded images to

--data/images

Run

python data_generator.py
python data_splitter.py

to distort and split the fisheye dataset. The distorted fisheye images will be placed in

--data/distorted

and split fisheye images will be placed in

--data/train_input
--data/train_target
--data/test_input
--data/test_target

Train

python train.py

Test

python test.py

Rectification

You can use your fisheye image. Before Start, make sure that the fisheye image have been placed in

--sample

Run

python calib.py

After rectification, the results will be placed in

--results

Demo

Additionally, you can also use the model using Gradio.

python web.py