Issue #3: StereoBM disparity width reduced problem (SSE2)(Resolved)

[aalajaji2012]

System information (version)

OpenCV => 3.1 (latest commit in master 3d57006) Operating System / Platform => Ubuntu 16.04.1 LTS Compiler => gcc (Ubuntu 5.4.0-6ubuntu1~16.04.2) Detailed description

Greetings,

I stumbled upon a problem while computing and analysing the disparity map created by StereoBM using SSE2.

For input images I used aloeL.jpg and aloeR.jpg from samples/data of OpenCV.

When computing a disparity map with cv::Mat_ as the type of the resulting disparity and with setPreFilterCap > 31 or setBlockSize > 21, I get a normal looking disparity map:

Parameters used:

disparities: 160 block size: 21 pre filter cap: 32 Scale of the resulting map:

max: 159.059 min: -1 Result (scaled [0, 255]): aloe_disp_cap_32

https://cloud.githubusercontent.com/assets/4293865/17890321/80ba5a3e-6935-11e6-812b-e4a9e35af25a.png

The problem arises when I change setPreFilterCap to 31. The horizontal resolution of the disparity map is reduced and the scale of the resulting map jumped by roughly 4096 times.

Parameters used:

disparities: 160 block size: 21 pre filter cap: 31 Scale of the resulting map:

max: 651520 min: -4096 Result (scaled [0, 255]): aloe_disp_cap_31 https://cloud.githubusercontent.com/assets/4293865/17890539/f5e7bfd0-6936-11e6-95ec-0c4a919fe64c.png When I traced the cause of this problem I found this line to be the root of the problem:

modules/calib3d/src/stereobm.cpp [line 1170-1172]:

if CV_SSE2

    bool useShorts = params.preFilterCap <= 31 && params.SADWindowSize <= 21 && checkHardwareSupport(CV_CPU_SSE2);

else

[Edit 1: modules/calib3d/src/stereobm.cpp [line 1016-1021]:

if CV_SSE2

    if( useShorts )
        findStereoCorrespondenceBM_SSE2( left_i, right_i, disp_i, cost_i, *state, ptr, row0, rows - row1 );
    else

endif

        findStereoCorrespondenceBM( left_i, right_i, disp_i, cost_i, *state, ptr, row0, rows - row1 );

The condition of useShorts matches with my findings, so if useShorts is true the problem arises if not, all is good. So it seems the SSE2 implementation is broken. Edit 1]

On the other hand if the type of resulting map is declared as cv::Mat the map's horizontal resolution is not reduced but the scale jumps by 16 times.

Parameters used:

disparities: 160 block size: 21 pre filter cap: 31 Scale of the resulting map:

max: 2544 min: -16 Result (scaled [0, 255]): aloe_disp_cap_31 https://cloud.githubusercontent.com/assets/4293865/17890902/1adf5634-6939-11e6-89bb-27d909f47285.png IMHO this is not the supposed behaviour because changing the pre-filter parameters should not affect the width and/or scale of the disparity map.

Cheers

Steps to reproduce

`#include

include <opencv2/calib3d.hpp>

include <opencv2/highgui.hpp>

int main(int argc, char* argv[]) {

cv::Mat leftFrame, rightFrame;

leftFrame = cv::imread(argv[1], cv::IMREAD_GRAYSCALE);
rightFrame = cv::imread(argv[2], cv::IMREAD_GRAYSCALE);

int numDisparities = 160;

cv::Ptr<cv::StereoBM> sbm;
sbm = cv::StereoBM::create(numDisparities, 21);

sbm->setPreFilterCap(31);

cv::Mat_<float> disp;
cv::Mat disp8;

sbm->compute(leftFrame, rightFrame, disp);

double min, max;
cv::minMaxLoc(disp, &min, &max);
std::cout << "Max: " << max << " Min: " << min << std::endl;

disp.convertTo(disp8, CV_8UC1, 255. / (numDisparities - 1));
cv::imshow("Disparity", disp8);
cv::waitKey(0);

return 0;

}`

[bedbad]

The issue is solved and it happened to be not connected to filters but to Mat_ converting ot Mat at some point of sbm function. It happens then rounding is happened all resulted coordinates halphed((1.99-> (1.00)). Solution: this filter works with Mat (integers) just fine. The result behaviour is expected.