modified focal estimation function in opencv_stitching

modules/stitching/autocalib.cpp

@@ -8,99 +8,63 @@ void focalsFromHomography(const Mat& H, double &f0, double &f1, bool &f0_ok, boo
 {
     CV_Assert(H.type() == CV_64F && H.size() == Size(3, 3));
 
-    const double h[9] =
-    {
-        H.at<double>(0, 0), H.at<double>(0, 1), H.at<double>(0, 2),
-        H.at<double>(1, 0), H.at<double>(1, 1), H.at<double>(1, 2),
-        H.at<double>(2, 0), H.at<double>(2, 1), H.at<double>(2, 2)
-    };
+    const double* h = reinterpret_cast<const double*>(H.data);
+
+    double d1, d2; // Denominators
+    double v1, v2; // Focal squares value candidates
 
     f1_ok = true;
-    double denom1 = h[6] * h[7];
-    double denom2 = (h[7] - h[6]) * (h[7] + h[6]);
-    if (max(abs(denom1), abs(denom2)) < 1e-5)
-        f1_ok = false;
-    else
-    {
-        double val1 = -(h[0] * h[1] + h[3] * h[4]) / denom1;
-        double val2 = (h[0] * h[0] + h[3] * h[3] - h[1] * h[1] - h[4] * h[4]) / denom2;
-        if (val1 < val2)
-            swap(val1, val2);
-        if (val1 > 0 && val2 > 0)
-            f1 = sqrt(abs(denom1) > abs(denom2) ? val1 : val2);
-        else if (val1 > 0)
-            f1 = sqrt(val1);
-        else
-            f1_ok = false;
-    }
+    d1 = h[6] * h[7];
+    d2 = (h[7] - h[6]) * (h[7] + h[6]);
+    v1 = -(h[0] * h[1] + h[3] * h[4]) / d1;
+    v2 = (h[0] * h[0] + h[3] * h[3] - h[1] * h[1] - h[4] * h[4]) / d2;
+    if (v1 < v2) swap(v1, v2);
+    if (v1 > 0 && v2 > 0) f1 = sqrt(abs(d1) > abs(d2) ? v1 : v2);
+    else if (v1 > 0) f1 = sqrt(v1);
+    else f1_ok = false;
 
     f0_ok = true;
-    denom1 = h[0] * h[3] + h[1] * h[4];
-    denom2 = h[0] * h[0] + h[1] * h[1] - h[3] * h[3] - h[4] * h[4];
-    if (max(abs(denom1), abs(denom2)) < 1e-5)
-        f0_ok = false;
-    else
-    {
-        double val1 = -h[2] * h[5] / denom1;
-        double val2 = (h[5] * h[5] - h[2] * h[2]) / denom2;
-        if (val1 < val2)
-            swap(val1, val2);
-        if (val1 > 0 && val2 > 0)
-            f0 = sqrt(abs(denom1) > abs(denom2) ? val1 : val2);
-        else if (val1 > 0)
-            f0 = sqrt(val1);
-        else
-            f0_ok = false;
-    }
+    d1 = h[0] * h[3] + h[1] * h[4];
+    d2 = h[0] * h[0] + h[1] * h[1] - h[3] * h[3] - h[4] * h[4];
+    v1 = -h[2] * h[5] / d1;
+    v2 = (h[5] * h[5] - h[2] * h[2]) / d2;
+    if (v1 < v2) swap(v1, v2);
+    if (v1 > 0 && v2 > 0) f0 = sqrt(abs(d1) > abs(d2) ? v1 : v2);
+    else if (v1 > 0) f0 = sqrt(v1);
+    else f0_ok = false;
 }
 
 
-bool focalsFromFundamental(const Mat &F, double &f0, double &f1)
+double estimateFocal(const vector<Mat> &images, const vector<ImageFeatures> &/*features*/, 
+                     const vector<MatchesInfo> &pairwise_matches)
 {
-    CV_Assert(F.type() == CV_64F);
-    CV_Assert(F.size() == Size(3, 3));
-
-    Mat Ft = F.t();
-    Mat k = Mat::zeros(3, 1, CV_64F);
-    k.at<double>(2, 0) = 1.f;
-
-    // 1. Compute quantities
-    double a = normL2sq(F*Ft*k) / normL2sq(Ft*k);
-    double b = normL2sq(Ft*F*k) / normL2sq(F*k);
-    double c = sqr(k.dot(F*k)) / (normL2sq(Ft*k) * normL2sq(F*k));
-    double d = k.dot(F*Ft*F*k) / k.dot(F*k);
-    double A = 1/c + a - 2*d;
-    double B = 1/c + b - 2*d;
-    double P = 2*(1/c - 2*d + 0.5*normL2sq(F));
-    double Q = -(A + B)/c + 0.5*(normL2sq(F*Ft) - 0.5*sqr(normL2sq(F)));
+    const int num_images = static_cast<int>(images.size());
 
-    // 2. Solve quadratic equation Z*Z*a_ + Z*b_ + c_ = 0
-    double a_ = 1 + c*P;
-    double b_ = -(c*P*P + 2*P + 4*c*Q);
-    double c_ = P*P + 4*c*P*Q + 12*A*B;
-    double D = b_*b_ - 4*a_*c_;
-    if (abs(D) < 1e-5)
-        D = 0;
-    else if (D < 0)
-        return false;
-    double D_sqrt = sqrt(D);
-    double Z0 = (-b_ - D_sqrt) / (2*a_);
-    double Z1 = (-b_ + D_sqrt) / (2*a_);
-
-    // 3. Choose solution
-    double w0 = abs(Z0*Z0*Z0 - 3*P*Z0*Z0 + 2*(P*P + 2*Q)*Z0 - 4*(P*Q + 4*A*B/c));
-    double w1 = abs(Z1*Z1*Z1 - 3*P*Z1*Z1 + 2*(P*P + 2*Q)*Z1 - 4*(P*Q + 4*A*B/c));
-    double Z = Z0;
-    if (w1 < w0)
-        Z = Z1;
+    vector<double> focals;
+    for (int src_idx = 0; src_idx < num_images; ++src_idx)
+    {
+        for (int dst_idx = 0; dst_idx < num_images; ++dst_idx)
+        {
+            const MatchesInfo &m = pairwise_matches[src_idx*num_images + dst_idx];
+            if (m.H.empty())
+                continue;
 
-    // 4.
-    double X = -1/c*(1 + 2*B/(Z - P));
-    double Y = -1/c*(1 + 2*A/(Z - P));
+            double f0, f1;
+            bool f0ok, f1ok;
+            focalsFromHomography(m.H, f0, f1, f0ok, f1ok);
+            if (f0ok && f1ok) focals.push_back(sqrt(f0*f1));
+        }
+    }
 
-    // 5. Compute focal lengths
-    f0 = 1/sqrt(1 + X/normL2sq(Ft*k));
-    f1 = 1/sqrt(1 + Y/normL2sq(F*k));
+    if (focals.size() + 1 >= images.size())
+    {
+        nth_element(focals.begin(), focals.end(), focals.begin() + focals.size()/2);
+        return focals[focals.size()/2];
+    }
 
-    return true;
+    LOGLN("Can't estimate focal length, will use naive approach");
+    double focals_sum = 0;
+    for (int i = 0; i < num_images; ++i)
+        focals_sum += images[i].rows + images[i].cols;
+    return focals_sum / num_images;
 }

modules/stitching/autocalib.hpp

 #ifndef __OPENCV_AUTOCALIB_HPP__
 #define __OPENCV_AUTOCALIB_HPP__
 
+#include <vector>
 #include <opencv2/core/core.hpp>
+#include "matchers.hpp"
 
 // See "Construction of Panoramic Image Mosaics with Global and Local Alignment"
 // by Heung-Yeung Shum and Richard Szeliski.
 void focalsFromHomography(const cv::Mat &H, double &f0, double &f1, bool &f0_ok, bool &f1_ok);
 
-bool focalsFromFundamental(const cv::Mat &F, double &f0, double &f1);
+double estimateFocal(const std::vector<cv::Mat> &images, const std::vector<ImageFeatures> &features, 
+                     const std::vector<MatchesInfo> &pairwise_matches);
 
 #endif // __OPENCV_AUTOCALIB_HPP__

modules/stitching/motion_estimators.cpp

@@ -64,46 +64,16 @@ struct CalcRotation
 };
 
 
-void HomographyBasedEstimator::estimate(const vector<Mat> &images, const vector<ImageFeatures> &/*features*/,
+void HomographyBasedEstimator::estimate(const vector<Mat> &images, const vector<ImageFeatures> &features,
                                         const vector<MatchesInfo> &pairwise_matches, vector<CameraParams> &cameras)
 {
     const int num_images = static_cast<int>(images.size());
 
-    // Find focals from pair-wise homographies
-    vector<bool> is_focal_estimated(num_images, false);
-    vector<double> focals;
-    for (int i = 0; i < num_images; ++i)
-    {
-        for (int j = 0; j < num_images; ++j)
-        {
-            int pair_idx = i * num_images + j;
-            if (pairwise_matches[pair_idx].H.empty())
-                continue;
-
-            double f_to, f_from;
-            bool f_to_ok, f_from_ok;
-            focalsFromHomography(pairwise_matches[pair_idx].H.inv(), f_to, f_from, f_to_ok, f_from_ok);
-
-            if (f_from_ok) focals.push_back(f_from);
-            if (f_to_ok) focals.push_back(f_to);
-
-            if (f_from_ok && f_to_ok)
-            {
-                is_focal_estimated[i] = true;
-                is_focal_estimated[j] = true;
-            }
-        }
-    }
-
-    is_focals_estimated_ = true;
-    for (int i = 0; i < num_images; ++i)
-        is_focals_estimated_ = is_focals_estimated_ && is_focal_estimated[i];
-
-    // Find focal median and use it as true focal length
-    nth_element(focals.begin(), focals.end(), focals.begin() + focals.size() / 2);
+    // Estimate focal length and set it for all cameras
+    double focal = estimateFocal(images, features, pairwise_matches);
     cameras.resize(num_images);
     for (int i = 0; i < num_images; ++i)
-        cameras[i].focal = focals[focals.size() / 2];
+        cameras[i].focal = focal;
 
     // Restore global motion
     Graph span_tree;