Using OpenCV to generate 3d points (assuming front

Question

I am currently trying to generate 3D points given stereo image pair in OpenCV. This has been done quite a bit as far as I can search.

I know the extrinsic parameters of the stereo setup which I'm going to assume is in frontal parallel configuration (really, it isn't that bad!). I know the focal length, baseline, and I'm going to assume the principal point as the center of the image (I know, I know...).

I calculate a psuedo-decent disparity map using StereoSGBM and hand coded the Q matrix following O'Reilly's Learning OpenCV book which specifies:

Q = [ 1 0 0      -c_x
      0 1 0      -c_y
      0 0 0      f
      0 0 -1/T_x (c_x - c_x')/T_x ]

I'll take that ( c_x, c_y ) is the principal point (which I specified in image coordinates), f is the focal length (which I described in mm), and T_x is the translation between the two cameras or baseline (which I also described in mm).

int type = CV_STEREO_BM_BASIC;
double rescx = 0.25, rescy = 0.25;
Mat disparity, vdisparity, depthMap;

Mat frame1 = imread( "C:\\Users\\Administrator\\Desktop\\Flow\\IMG137.jpg", CV_LOAD_IMAGE_GRAYSCALE );
Mat frame1L = frame1( Range( 0, frame1.rows ), Range( 0, frame1.cols/2 ));
Mat frame1R = frame1( Range( 0, frame1.rows ), Range( frame1.cols/2, frame1.cols ));

resize( frame1L, frame1L, Size(), rescx, rescy );
resize( frame1R, frame1R, Size(), rescx, rescy );

int preFilterSize = 9, preFilterCap = 32, disparityRange = 4;
int minDisparity = 2, textureThreshold = 12, uniquenessRatio = 3;
int windowSize = 21, smoothP1 = 0, smoothP2 = 0, dispMaxDiff = 32;
int speckleRange = 0, speckleWindowSize = 0;

bool dynamicP = false;

StereoSGBM stereo( minDisparity*-16, disparityRange*16, windowSize,
    smoothP1, smoothP2, dispMaxDiff,
    preFilterCap, uniquenessRatio,
    speckleRange*16, speckleWindowSize, dynamicP );

stereo( frame1L, frame1R, disparity );

double m1[3][3] = { { 46, 0, frame1L.cols/2 }, { 0, 46, frame1L.rows/2 }, { 0, 0, 1 } };
double t1[3] = { 65, 0, 0 };
double q[4][4] = {{ 1, 0, 0, -frame1L.cols/2.0 }, { 0, 1, 0, -frame1L.rows/2.0 }, { 0, 0, 0, 46 }, { 0, 0, -1.0/65, 0 }};
Mat cm1( 3, 3, CV_64F, m1), cm2( 3, 3, CV_64F, m1), T( 3, 1, CV_64F, t1 );
Mat R1, R2, P1, P2;
Mat Q( 4, 4, CV_64F, q );

//stereoRectify( cm1, Mat::zeros( 5, 1, CV_64F ), cm2, Mat::zeros( 5, 1, CV_64F ),  frame1L.size(), Mat::eye( 3, 3, CV_64F ), T, R1, R2, P1, P2, Q ); 

normalize( disparity, vdisparity, 0, 256, NORM_MINMAX );
//convertScaleAbs( disparity, disparity, 1/16.0 );
reprojectImageTo3D( disparity, depthMap, Q, true );
imshow( "Disparity", vdisparity );
imshow( "3D", depthMap );

So I feed the resulting disparity map from StereoSGBM and that Q matrix to get 3D points, which I write out to a ply file.

But the result is this: http://i.stack.imgur.com/7eH9V.png

Fun to look at, but not what I need :(. I read online that it gets better results after dividing the disparity map by 16 and indeed it looked marginally better (it actually looks like there was a camera that took the shot!).

This is my disparity map if you're interested: http://i.stack.imgur.com/lNPkO.png

I understand that without callibration, it's hardly going to look like the best 3d projection, but I was expecting something a bit... better.

Any suggestions?

Answer 1

Under fronto-parrallel assumption, the relation between disparity and 3D depth is: d = f*T/Z, where d is the disparity, f is the focal length, T is the baseline and Z is the 3D depth. If you treat the image center as the principal point, the 3D coordinate system is settled. Then for a pixel (px,py), its 3D coordinate (X, Y, Z) is:

X = (px-cx)*Z/f, Y = (py- cy)*Z/f, Z = f*T/d,

where cx, cy are the pixel coordinate of image center.

Your disparity image seems pretty good and it can generate reasonable 3D point clouds.

A simple disparity browser on github.