Main Content

Structure from Motion from Multiple Views

Structure from motion (SfM) is the process of estimating the 3-D structure of a scene from a set of 2-D views. It is used in many applications, such as robot navigation, autonomous driving, and augmented reality. This example shows you how to estimate the poses of a calibrated camera from a sequence of views, and reconstruct the 3-D structure of the scene up to an unknown scale factor.

Overview

This example shows how to reconstruct a 3-D scene from a sequence of 2-D views taken with a camera calibrated using the Camera Calibrator. The example uses an imageviewset object to store and manage the data associated with each view, such as the camera pose and the image points, as well as matches between points from pairs of views.

The example uses the pairwise point matches to estimate the camera pose of the current view relative to the previous view. It then links the pairwise matches into longer point tracks spanning multiple views using the findTracks method of the imageviewset object. These tracks then serve as inputs to multiview triangulation using the triangulateMultiview function and the refinement of camera poses and the 3-D scene points using the bundleAdjustment function.

The example consists of two main parts: camera motion estimation and dense scene reconstruction. In the first part, the example estimates the camera pose for each view using a sparse set of points matched across the views. In the second part, the example iterates over the sequence of views again, using vision.PointTrackerto track a dense set of points across the views, to compute a dense 3-D reconstruction of the scene.

The camera motion estimation algorithm consists of the following steps:

  1. For each pair of consecutive images, find a set of point correspondences. This example detects the interest points using the detectSURFFeatures function, extracts the feature descriptors using the extractFeatures functions, and finds the matches using the matchFeatures function. Alternatively, you can track the points across the views using vision.PointTracker.

  2. Estimate the relative pose of the current view, which is the camera orientation and location relative to the previous view. The example uses a helper function helperEstimateRelativePose, which calls estimateEssentialMatrix and estrelpose.

  3. Transform the relative pose of the current view into the coordinate system of the first view of the sequence.

  4. Store the current view attributes: the camera pose and the image points.

  5. Store the inlier matches between the previous and the current view.

  6. Find point tracks across all the views processed so far.

  7. Use the triangulateMultiview function to compute the initial 3-D locations corresponding to the tracks.

  8. Use the bundleAdjustment function to refine the camera poses and the 3-D points. Store the refined camera poses in the imageviewset object.

Read the Input Image Sequence

Read and display the image sequence.

% Use |imageDatastore| to get a list of all image file names in a directory.
imageDir = fullfile(toolboxdir('vision'), 'visiondata', ...
      'structureFromMotion');
imds = imageDatastore(imageDir);

% Display the images.
figure
montage(imds.Files, 'Size', [3, 2]);

% Convert the images to grayscale.
images = cell(1, numel(imds.Files));
for i = 1:numel(imds.Files)
    I = readimage(imds, i);
    images{i} = im2gray(I);
end

title('Input Image Sequence');

Load Camera Parameters

Load the cameraParameters object created using the Camera Calibrator.

data = load(fullfile(imageDir, 'cameraParams.mat'));
cameraParams = data.cameraParams;

Create a View Set Containing the First View

Use an imageviewset object to store and manage the image points and the camera pose associated with each view, as well as point matches between pairs of views. Once you populate an imageviewset object, you can use it to find point tracks across multiple views and retrieve the camera poses to be used by triangulateMultiview and bundleAdjustment functions.

% Get intrinsic parameters of the camera
intrinsics = cameraParams.Intrinsics;

% Undistort the first image.
I = undistortImage(images{1}, intrinsics); 

% Detect features. Increasing 'NumOctaves' helps detect large-scale
% features in high-resolution images. Use an ROI to eliminate spurious
% features around the edges of the image.
border = 50;
roi = [border, border, size(I, 2)- 2*border, size(I, 1)- 2*border];
prevPoints   = detectSURFFeatures(I, NumOctaves=8, ROI=roi);

% Extract features. Using 'Upright' features improves matching, as long as
% the camera motion involves little or no in-plane rotation.
prevFeatures = extractFeatures(I, prevPoints, Upright=true);

% Create an empty imageviewset object to manage the data associated with each
% view.
vSet = imageviewset;

% Add the first view. Place the camera associated with the first view
% and the origin, oriented along the Z-axis.
viewId = 1;
vSet = addView(vSet, viewId, rigidtform3d, Points=prevPoints);

Add the Rest of the Views

Go through the rest of the images. For each image:

  1. Match points between the previous and the current image.

  2. Estimate the camera pose of the current view relative to the previous view.

  3. Compute the camera pose of the current view in the global coordinate system relative to the first view.

  4. Triangulate the initial 3-D world points.

  5. Use bundle adjustment to refine all camera poses and the 3-D world points.

for i = 2:numel(images)
    % Undistort the current image.
    I = undistortImage(images{i}, intrinsics);
    
    % Detect, extract and match features.
    currPoints   = detectSURFFeatures(I, NumOctaves=8, ROI=roi);
    currFeatures = extractFeatures(I, currPoints, Upright=true);    
    indexPairs   = matchFeatures(prevFeatures, currFeatures, ...
        MaxRatio=0.7, Unique=true);
    
    % Select matched points.
    matchedPoints1 = prevPoints(indexPairs(:, 1));
    matchedPoints2 = currPoints(indexPairs(:, 2));
    
    % Estimate the camera pose of current view relative to the previous view.
    % The pose is computed up to scale, meaning that the distance between
    % the cameras in the previous view and the current view is set to 1.
    % This will be corrected by the bundle adjustment.
    [relPose, inlierIdx] = helperEstimateRelativePose(...
        matchedPoints1, matchedPoints2, intrinsics);
    
    % Get the table containing the previous camera pose.
    prevPose = poses(vSet, i-1).AbsolutePose;
        
    % Compute the current camera pose in the global coordinate system 
    % relative to the first view.
    currPose = rigidtform3d(prevPose.A*relPose.A);
    
    % Add the current view to the view set.
    vSet = addView(vSet, i, currPose, Points=currPoints);

    % Store the point matches between the previous and the current views.
    vSet = addConnection(vSet, i-1, i, relPose, Matches=indexPairs(inlierIdx,:));
    
    % Find point tracks across all views.
    tracks = findTracks(vSet);

    % Get the table containing camera poses for all views.
    camPoses = poses(vSet);

    % Triangulate initial locations for the 3-D world points.
    xyzPoints = triangulateMultiview(tracks, camPoses, intrinsics);
    
    % Refine the 3-D world points and camera poses.
    [xyzPoints, camPoses, reprojectionErrors] = bundleAdjustment(xyzPoints, ...
        tracks, camPoses, intrinsics, FixedViewId=1, ...
        PointsUndistorted=true);

    % Store the refined camera poses.
    vSet = updateView(vSet, camPoses);

    prevFeatures = currFeatures;
    prevPoints   = currPoints;  
end

Display Camera Poses

Display the refined camera poses and 3-D world points.

% Display camera poses.
camPoses = poses(vSet);
figure;
plotCamera(camPoses, Size=0.2);
hold on

% Exclude noisy 3-D points.
goodIdx = (reprojectionErrors < 5);
xyzPoints = xyzPoints(goodIdx, :);

% Display the 3-D points.
pcshow(xyzPoints, VerticalAxis='y', VerticalAxisDir='down', MarkerSize= 45);
grid on
hold off

% Specify the viewing volume.
loc1 = camPoses.AbsolutePose(1).Translation;
xlim([loc1(1)-5, loc1(1)+4]);
ylim([loc1(2)-5, loc1(2)+4]);
zlim([loc1(3)-1, loc1(3)+20]);
camorbit(0, -30);

title('Refined Camera Poses');

Compute Dense Reconstruction

Go through the images again. This time detect a dense set of corners, and track them across all views usingvision.PointTracker.

% Read and undistort the first image
I = undistortImage(images{1}, intrinsics); 

% Detect corners in the first image.
prevPoints = detectMinEigenFeatures(I, MinQuality=0.001);

% Create the point tracker object to track the points across views.
tracker = vision.PointTracker(MaxBidirectionalError=1, NumPyramidLevels=6);

% Initialize the point tracker.
prevPoints = prevPoints.Location;
initialize(tracker, prevPoints, I);

% Store the dense points in the view set.

vSet = updateConnection(vSet, 1, 2, Matches=zeros(0, 2));
vSet = updateView(vSet, 1, Points=prevPoints);

% Track the points across all views.
for i = 2:numel(images)
    % Read and undistort the current image.
    I = undistortImage(images{i}, intrinsics); 
    
    % Track the points.
    [currPoints, validIdx] = step(tracker, I);
    
    % Clear the old matches between the points.
    if i < numel(images)
        vSet = updateConnection(vSet, i, i+1, Matches=zeros(0, 2));
    end
    vSet = updateView(vSet, i, Points=currPoints);
    
    % Store the point matches in the view set.
    matches = repmat((1:size(prevPoints, 1))', [1, 2]);
    matches = matches(validIdx, :);        
    vSet = updateConnection(vSet, i-1, i, Matches=matches);
end

% Find point tracks across all views.
tracks = findTracks(vSet);

% Find point tracks across all views.
camPoses = poses(vSet);

% Triangulate initial locations for the 3-D world points.
xyzPoints = triangulateMultiview(tracks, camPoses,...
    intrinsics);

% Refine the 3-D world points and camera poses.
[xyzPoints, camPoses, reprojectionErrors] = bundleAdjustment(...
    xyzPoints, tracks, camPoses, intrinsics, FixedViewId=1, ...
    PointsUndistorted=true);

Display Dense Reconstruction

Display the camera poses and the dense point cloud.

% Display the refined camera poses.
figure;
plotCamera(camPoses, Size=0.2);
hold on

% Exclude noisy 3-D world points.
goodIdx = (reprojectionErrors < 5);

% Display the dense 3-D world points.
pcshow(xyzPoints(goodIdx, :), VerticalAxis='y', VerticalAxisDir='down', MarkerSize=45);
grid on
hold off

% Specify the viewing volume.
loc1 = camPoses.AbsolutePose(1).Translation;
xlim([loc1(1)-5, loc1(1)+4]);
ylim([loc1(2)-5, loc1(2)+4]);
zlim([loc1(3)-1, loc1(3)+20]);
camorbit(0, -30);

title('Dense Reconstruction');

References

[1] M.I.A. Lourakis and A.A. Argyros (2009). "SBA: A Software Package for Generic Sparse Bundle Adjustment". ACM Transactions on Mathematical Software (ACM) 36 (1): 1-30.

[2] R. Hartley, A. Zisserman, "Multiple View Geometry in Computer Vision," Cambridge University Press, 2003.

[3] B. Triggs; P. McLauchlan; R. Hartley; A. Fitzgibbon (1999). "Bundle Adjustment: A Modern Synthesis". Proceedings of the International Workshop on Vision Algorithms. Springer-Verlag. pp. 298-372.