gpu-based-image-stitching

gpu-based-image-stitching copied to clipboard

0x00 Intro

This project is a simple implementation of opencv for the following papers.

Du, Chengyao, et al. (2020). GPU based parallel optimization for real time panoramic video stitching. Pattern Recognition Letters, 133, 62-69.

Fast panorama stitching method using UMat.

Speed of 4 cameras at 4k resolution is greater than 200fps in 1080ti.

This project does not provide a dataset so it cannot be used out of the box.

一个使用 OpenCV 进行快速全景视频拼接的方法。通过巧妙的流并行策略，在 1080ti 上可以对 4k 视频进行超过 200fps 的图像拼接。

0x01 Quick Start

$ mkdir build && cd build
$ cmake ..
$ make
$ ./image-stitching

0x02 Known Issue

• Real FPS is much slower than FPS: imwrite takes a lot of time.
• Stitching will never stop: I didn't think about the way to use the dataset before, so I added a dataset branch. You can follow this commit to modify your code. It will stop after stitching, but the program will not exit. The dataset datasets/air-4cam-mp4 will generate 149 pictures.

Welcome to pull requests!

0x03 Example

About these procedure below (chinese) http://s1nh.com/post/image-stitching-post-process/ .

00.mp4	01.mp4	02.mp4	03.mp4

stitching

exposure-mask

exposure-mask-refine

apply-mask

final-panorama