Stereo matching through data parallel c++

This project is an example of stereo matching algorithm written based on OneAPI dpc++, including two kernel types, Basic and ND-Range for GPU.

The cost function of stereo matching is the simple Sum of Absolute Difference (SAD) of the local window . In order to speed up the calculation, the original C++ code has converted the two-dimensional cost calculation into: first calculate the one-dimensional cost in the vertical direction, and then aggregate in the horizontal direction. The disparity position is calculated using the Winner Takes All (WTA) algorithm.

This project has been tested on Windows system and Intel's DevCloud environment.

Usage instructions:

1. Windows environment：

Windows 10 Professional Edition, 64-bit, Intel Core i7-7700 CPU, HD Graphics 630; it can run successfully in Visual Studio 2017 Release x64 mode, but there will be problems in Debug mode. After running successfully, it will display as follows:

Filename: ../res/imL.png W: 1280 H: 872

Running on Intel(R) Graphics [0x5912]
Max WG size is 256
Max Global memory size is 6830936064
Max Local memory size is 65536

Evaluating ND-Range Kernel...

Start image processing with offloading to GPU...
--The processing time is 0.77529 seconds
Start image processing with offloading to GPU...
--The processing time is 0.590326 seconds
Start image processing with offloading to GPU...
--The processing time is 0.604187 seconds
Start image processing with offloading to GPU...
--The processing time is 0.600635 seconds
Start image processing with offloading to GPU...
--The processing time is 0.600033 seconds

Average time for image processing:
--The average processing time was 0.634094 seconds

Evaluating Basic Kernel...

Start image processing with offloading to GPU...
--The processing time is 1.30688 seconds
Start image processing with offloading to GPU...
--The processing time is 1.25801 seconds
Start image processing with offloading to GPU...
--The processing time is 1.2814 seconds
Start image processing with offloading to GPU...
--The processing time is 1.317 seconds
Start image processing with offloading to GPU...
--The processing time is 1.27447 seconds

Average time for image processing:
--The average processing time was 1.28755 seconds


/*********Compare Result*********/
All results Matched!
--The CPU processing time is 1.71865 seconds

Stereo matching successfully completed on the device.
The processed image has been written to ../res/disparity.png

The calculated disparity map is shown as follows:

2. Linux system：

(1) Set environment variables

source /opt/intel/oneapi/setvars.sh

(2) Copy the code repository

git clone https://github.com/silverfly1992/stereo-matching-dpc.git

(3) Compile and run

cd stereo-matching-dpc rm -rf build mkdir build && cd build && cmake ../. && make make run

If you want to verify the project on Intel's DevCloud, just execute run_scripts.ipynb. The screenshot of the execution result is as follows:

Inspiration

This application was inspired by the The Great Cross Architecture Challenge.

This project referred to the CUDA code in https://github.com/dhernandez0/sgm.