Frame of 42yeah

Let's perform 3D reconstruction from multiple images! Day 2


Today, we are going to setup the project!

As performing SIFT/ORB itself is way too easy in OpenCV (there's literally a call for that), we are going to skip that and begin our very first big module: Structure from Motion (SfM).

Input Data

Before that though, we need a dataset. I am using my own dataset here, a statue of the renowned physicist Yang Chen-Ning in front of my school library. You can get it here. Or, if you are not a big fan of him (who isn't?), you can use royshil/SfM-Toy-Library's Crazyhorse dataset. Or, you can just make it yourself. Just take photos of the same object at slightly varying angles. After that, we can finally begin.


The Main Program

First, we are going to write a main.cpp:

#include <iostream>
#include <opencv2/imgcodecs.hpp>
#include "SfM.hpp"

int main(int argc, const char * argv[]) {
    SfM sfm("assets");;
    return 0;

SfM Header

Easy, right? So, time to define SfM!

#include <vector>
#include <string>
#include <opencv2/imgcodecs.hpp>

#define LOG(fmt, ...) printf(fmt "\n", ##__VA_ARGS__)

typedef std::vector<cv::Mat> Images;

class SfM {
    SfM() {}

    SfM(std::string path) {
    void run();

    void set_input_path(std::string path);

    Images views;

For the sake of debugging I've added quite a bit of stuffs here.

SfM Implementation (For Now)

void SfM::set_input_path(std::string path) {
    for (const auto &entry : std::filesystem::directory_iterator(path)) {
        LOG("Adding %s to views...", entry.path().c_str());

void SfM::run() { 
    // TODO: Run SfM. After this, we will get a 3D pointcloud. Yay!

Well, that's it! All that it does now is load cv::Mat into an array. That will soon change, though!

Feature Extraction

Feature Header

As now we have an array of images, we really should get their features, amirite? So first, we need to define feature:

struct Feature {
    std::vector<cv::KeyPoint> keypoints;
    std::vector<cv::Point2f> points;
    cv::Mat descriptor;

typedef std::vector<Feature> Features;

And then we should pop this into SfM class:

void detect_features();
Features features;

And add this into run():



After that, it's time to implement! We are going to use the ORB (Oriented FAST and Rotated BRIEF) detector to calculate keypoints & descriptors.

void SfM::detect_features() {
    cv::Ptr<cv::FeatureDetector> detector = cv::ORB::create(5000);
    for (int i = 0; i < views.size(); i++) {
        const cv::Mat &view = views[i];
        Feature feature;
        detector->detectAndCompute(view, cv::noArray(), feature.keypoints, feature.descriptor);
        for (int i = 0; i < feature.keypoints.size(); i++) {

That's it for keypoints. If you are curious and want to see its output, it should look a little something like this:


Feature Matching

After all features are extracted, now's the time to match them. As we have quite a few input views, we are going to match them, pair by pair, and put that into a matrix of sort. This matrix is going to be symmetric, of course, so we really only need half of them.

Modification to SfM Header

Simply pop this to your SfM header:

typedef std::vector<cv::DMatch > Matches;
std::vector<std::vector<Matches> > feature_matching_mat;

and we are ready to go.


Since this is still (kinda) feature detection, I am simply going to append the following code to detect_features() above. You can do it however you like.

cv::Ptr<cv::DescriptorMatcher> matcher = cv::DescriptorMatcher::create("BruteForce-Hamming");
for (int i = 0; i < views.size(); i++) {
    for (int j = i + 1; j < views.size(); j++) {
        std::vector<Matches> initial_matches;
        Matches pruned_matches;
        matcher->knnMatch(features[i].descriptor, features[j].descriptor, initial_matches, 2);

        for (int i = 0; i < initial_matches.size(); i++) {
            if (initial_matches[i][0].distance < 0.8f * initial_matches[i][1].distance) {

        feature_matching_mat[i][j] = pruned_matches;

Keep in mind that half of the matrix does not have any data at all:

Such empty

If you are curious for the matches, they look a little bit like this:

Sample matches

That's It For Today

...And we are gonna stop here for now. That's a lot to take in for today, but tomorrow is going to be even harder. But don't give up yet! Tomorrow, we are going to do the baseline triangulation, which will generate our primitive point cloud!