Skip to content

Latest commit

 

History

History

sgm_cuda_256

README.md

Semi-Global Matching on the GPU in the paper

Introduction

A GPU base SGM, use as a base-line method.The orginal code is clone from origin code. Because for the origin code, the disparity range is 128, and this is too small for the application, so here we will provide the code with a disparity range is 256.

If you want to use the origin code setting (depends on OpenCV), we will try provide it, the code can be found in folder WithOpenCV.

Update code for Cluster(CNES)

Because there is no OpenCV in the Cluster, and I do not want to install OpenCV (too big, and no sudo), so this code is modified with OpenCV files, so that the code doesn't depend on OpenCV.

On the other hand, because the compared methods have sub-pixel accuracy, the output file is 16bit, i.e. unsigned short.

Dependency library

  1. remove the OpenCV library, so all the needed OpenCV files are in folder.
  2. use png++, this is used to read and write file, so that this code only support PNG files, the input and output are both PNG file.

Disparity range

In the code, the PATH_AGGREGATION is 8 for defualt, and the MAX_DISPARITY is 256 now.

From the code, the WARP_SIZE is 32(fixed). So for 32 thread, run one piexl, so that the batch size is MAX_DISPARITY/WARP_SIZE. Considering that if the MAX_DISPARITY is 256, the batch size is 8. Shortage is that batch size is 4 is more clear. The Algorithm 3 is the main idea for the code, the parameters are shared in the 32 thread.

Compile

To compile the CUDA code, the compatibility is important, you can set the compile gcc/g++ before make, for my computer, the version is Cuda compilation tools, release 11.5, V11.5.119, so we can run like this :

export CC=/usr/bin/gcc-11
export CXX=/usr/bin/g++-11

mkdir build
cd build
cmake ..
make

Exmaple

we give an example in the folder for the Toulouse Métropole dataset.

./example.sh

TODO

  • check the label recompute

recompute means calculating the cost again, now recompute is not used in DIR_LEFTRIGHT and DIR_RIGHTLEFT. If so, the code can be much cleaner. (T *rp0, T *rp1, T *rp2, T *rp3, T *rp4, T *rp5, T *rp6, T *rp7) can be removed. The iterations in DIR_LEFTRIGHT and DIR_RIGHTLEFT can be removed.

  • make the code clean

  • sub-pixel accuracy

  • left-right check

Feed Back

If you think you have any problem, contact [Teng Wu]whuwuteng@gmail.com