Sparse 3D reconstruction MATLAB example - matlab

Question

I have a stereo camera system and I am trying this MATLAB's Computer Vision toolbox example (http://www.mathworks.com/help/vision/ug/sparse-3-d-reconstruction-from-multiple-views.html) with my own images and camera calibration files. I used Caltech's camera calibration toolbox (http://www.vision.caltech.edu/bouguetj/calib_doc/).
First I tried each camera separately based on first example and found the intrinsic camera calibration matrices for each camera and saved them. I also undistorted the left and right images using Caltech toolbox. Therefore I commented out the code for that from MATLAB example.
Here are the instrinsic camera matrices:
K1=[1050 0 630;0 1048 460;0 0 1];
K2=[1048 0 662;0 1047 468;0 0 1];
BTW, these are the right and center lenses from bumblebee XB3 cameras.
Question: aren't they supposed to be the same?
Then I did stereo calibration based on fifth example. I saved the rotation matrix (R) and translation matrix (T) from that. Therefore I commented out the code for that from MATLAB example.
Here are the rotation and translation matrices:
R=[0.9999 -0.0080 -0.0086;0.0080 1 0.0048;0.0086 -0.0049 1];
T=[120.14 0.55 1.04];
Then I fed all these images and calibration files and camera matrices to the MATLAB example and tried to find the 3-D point cloud but the results are not promising. I am attaching the code here. I think here are two problems:
1- My epipolar constraint value is too large!(to the power of 16)
2- I am not sure about the camera matrices and how I calculated them from R, and T from Caltech toolbox!
P.S. as far as feature extraction goes that is fine.
would be great if someone can help.
clear
close all
clc
files = {'Left1.tif';'Right1.tif'};
for i = 1:numel(files)
files{i}=fullfile('...\sparse_matlab', files{i});
images(i).image = imread(files{i});
end
figure;
montage(files); title('Pair of Original Images')
% Intrinsic camera parameters
load('Calib_Results_Left.mat')
K1 = KK;
load('Calib_Results_Right.mat')
K2 = KK;
%Extrinsics using stereo calibration
load('Calib_Results_stereo.mat')
Rotation=R;
Translation=T';
images(1).CameraMatrix=[Rotation; Translation] * K1;
images(2).CameraMatrix=[Rotation; Translation] * K2;
% Detect feature points and extract SURF descriptors in images
for i = 1:numel(images)
%detect SURF feature points
images(i).points = detectSURFFeatures(rgb2gray(images(i).image),...
'MetricThreshold',600);
%extract SURF descriptors
[images(i).featureVectors,images(i).points] = ...
extractFeatures(rgb2gray(images(i).image),images(i).points);
end
% Visualize several extracted SURF features from the Left image
figure; imshow(images(1).image);
title('1500 Strongest Feature Points from Globe01');
hold on;
plot(images(1).points.selectStrongest(1500));
indexPairs = ...
matchFeatures(images(1).featureVectors, images(2).featureVectors,...
'Prenormalized', true,'MaxRatio',0.4) ;
matchedPoints1 = images(1).points(indexPairs(:, 1));
matchedPoints2 = images(2).points(indexPairs(:, 2));
figure;
% Visualize correspondences
showMatchedFeatures(images(1).image,images(2).image,matchedPoints1,matchedPoints2,'montage' );
title('Original Matched Features from Globe01 and Globe02');
% Set a value near zero, It will be used to eliminate matches that
% correspond to points that do not lie on an epipolar line.
epipolarThreshold = .05;
for k = 1:length(matchedPoints1)
% Compute the fundamental matrix using the example helper function
% Evaluate the epipolar constraint
epipolarConstraint =[matchedPoints1.Location(k,:),1]...
*helperCameraMatricesToFMatrix(images(1).CameraMatrix,images(2).CameraMatrix)...
*[matchedPoints2.Location(k,:),1]';
%%%% here my epipolarConstraint results are bad %%%%%%%%%%%%%
% Only consider feature matches where the absolute value of the
% constraint expression is less than the threshold.
valid(k) = abs(epipolarConstraint) < epipolarThreshold;
end
validpts1 = images(1).points(indexPairs(valid, 1));
validpts2 = images(2).points(indexPairs(valid, 2));
figure;
showMatchedFeatures(images(1).image,images(2).image,validpts1,validpts2,'montage');
title('Matched Features After Applying Epipolar Constraint');
% convert image to double format for plotting
doubleimage = im2double(images(1).image);
points3D = ones(length(validpts1),4); % store homogeneous world coordinates
color = ones(length(validpts1),3); % store color information
% For all point correspondences
for i = 1:length(validpts1)
% For all image locations from a list of correspondences build an A
pointInImage1 = validpts1(i).Location;
pointInImage2 = validpts2(i).Location;
P1 = images(1).CameraMatrix'; % Transpose to match the convention in
P2 = images(2).CameraMatrix'; % in [1]
A = [
pointInImage1(1)*P1(3,:) - P1(1,:);...
pointInImage1(2)*P1(3,:) - P1(2,:);...
pointInImage2(1)*P2(3,:) - P2(1,:);...
pointInImage2(2)*P2(3,:) - P2(2,:)];
% Compute the 3-D location using the smallest singular value from the
% singular value decomposition of the matrix A
[~,~,V]=svd(A);
X = V(:,end);
X = X/X(end);
% Store location
points3D(i,:) = X';
% Store pixel color for visualization
y = round(pointInImage1(1));
x = round(pointInImage1(2));
color(i,:) = squeeze(doubleimage(x,y,:))';
end
% add green point representing the origin
points3D(end+1,:) = [0,0,0,1];
color(end+1,:) = [0,1,0];
% show images
figure('units','normalized','outerposition',[0 0 .5 .5])
subplot(1,2,1); montage(files,'Size',[1,2]); title('Original Images')
% plot point-cloud
hAxes = subplot(1,2,2); hold on; grid on;
scatter3(points3D(:,1),points3D(:,2),points3D(:,3),50,color,'fill')
xlabel('x-axis (mm)');ylabel('y-axis (mm)');zlabel('z-axis (mm)')
view(20,24);axis equal;axis vis3d
set(hAxes,'XAxisLocation','top','YAxisLocation','left',...
'ZDir','reverse','Ydir','reverse');
grid on
title('Reconstructed Point Cloud');

Answer 1

First of all, the Computer Vision System Toolbox now includes a Camera Calibrator App for calibrating a single camera, and also support for programmatic stereo camera calibration. It would be easier for you to use those tools, because the example you are using and the Caltech Calibration Toolbox use somewhat different conventions.
The example uses the pre-multiply convention, i.e. row vector * matrix, while the Caltech toolbox uses the post-multiply convention (matrix * column vector). That means that if you do use the camera parameters from Caltech, you would have to transpose the intrinsic matrix and the rotation matrices. That could be the main cause of your problems.
As far as the intrinsics being different between your two cameras, that is perfectly normal. All cameras are slightly different.
It would also help to see the matched features that you've used for triangulation. Given that you are reconstructing an elongated object, it doesn't seem too surprising to see the reconstructed points form a line in 3D...
You could also try rectifying the images and doing a dense reconstruction, as in the example I've linked to above.