A modified GPU-accelerated pm-stereo red/black modified version for 3d reconstruction.(Note: This is not original PM-STEREO algorithm)
For the idea, you could check gipuma paper and their project:
https://github.com/kysucix/gipuma
The algorithm was verified in a subset of WHU-Stereo dataset,you could download them here:
https://github.com/Sheng029/WHU-Stereo
The code is first written in Windows in 1080ti,and also test in 3060 with my laptop,but there's no extra change are needed to transport to Linux system.
CUDA 11.0 or above is required,the code is self-contained so you don't need any 3rd library.
In order to use std::filesystem,which is optional to find the path without dependency, you make want to set language standard to
"-std=c++17 -Xcompiler "/std:c++17"
You should compute a depth map in WHU-Stereo in 8-10s within 8 iteration, which have 100x-500x speed up than original Patch-Match Stereo CPU implementation.
There's a technical talk that a Row parallel modification of patch match stereo is actually used to generate real outdoor scene 3d information with satellite data using intel One API by Intel developers, the link with Chinese subtitle is here:
https://www.bilibili.com/video/BV1uh411b7Sj/
The Repo aim to update this algorithm soon.
UPDATE:THIS ALGO IS COMPLETE!You can choose in PMoption between PM_MODE::RB/PM_MODE::ROW_SWEEP to test different mode of algorithm.
Original Patch-Match Stereo:http://www.bmva.org/bmvc/2011/proceedings/paper14/paper14.pdf
RB-Modified Patch-Match for Multiview dense reconstruction:
Row-Parallel Patch-Match for Multiview dense reconstruction: