Brings back outlier detection based on the residuals

EyeRecToo/src/gaze-estimation/BinocularPolyFit.cpp

@@ -18,9 +18,6 @@ string BinocularPolyFit::description() const
 
 bool BinocularPolyFit::calibrate(vector<CollectionTuple>& calibrationTuples, QString& errorMsg)
 {
-    if (!PolyFit::calibrate(calibrationTuples, errorMsg))
-        return false;
-
     PointVector leftPupil;
     PointVector rightPupil;
     PointVector gaze;
@@ -39,9 +36,50 @@ bool BinocularPolyFit::calibrate(vector<CollectionTuple>& calibrationTuples, QSt
     binocularCalibrated = calibrate(plType, leftPupil.x, leftPupil.y, rightPupil.x, rightPupil.y, gaze.x, bcx, errorMsg)
         && calibrate(plType, leftPupil.x, leftPupil.y, rightPupil.x, rightPupil.y, gaze.y, bcy, errorMsg);
 
+    if (!binocularCalibrated)
+        return binocularCalibrated;
+
+    const auto& ref = calibrationTuples[0];
+    float threshold = 0.01f * hypot(ref.field.width, ref.field.height);
+    vector<CollectionTuple> inliers;
+    for (const auto& tuple : calibrationTuples) {
+        GazeEstimate l, r, b;
+        l.pupilConfidence = tuple.lEye.pupil.confidence;
+        r.pupilConfidence = tuple.rEye.pupil.confidence;
+        estimateBinocular2d(tuple, l, r, b);
+        Point2f gt = { tuple.field.collectionMarker.center.x, tuple.field.collectionMarker.center.y };
+        if (cv::norm(gt - b.gp) < threshold)
+            inliers.push_back(tuple);
+    }
+
+    binocularCalibrated = calibrateWithInliers(inliers, errorMsg);
     return binocularCalibrated;
 }
 
+bool BinocularPolyFit::calibrateWithInliers(std::vector<CollectionTuple>& calibrationTuples, QString& errorMsg)
+{
+    if (!PolyFit::calibrate(calibrationTuples, errorMsg))
+        return false;
+
+    PointVector leftPupil;
+    PointVector rightPupil;
+    PointVector gaze;
+
+    for (const auto& tuple : calibrationTuples) {
+        leftPupil.insert(PolyFit::normalize(tuple.lEye.pupil.center, tuple.lEye.input));
+        rightPupil.insert(PolyFit::normalize(tuple.rEye.pupil.center, tuple.rEye.input));
+        gaze.insert(PolyFit::normalize(tuple.field.collectionMarker.center, tuple.field.input));
+    }
+
+    if (gaze.size() < binocularUnknowns) {
+        errorMsg = QString("Not enough calibration points (%1/%2).").arg(gaze.size()).arg(binocularUnknowns);
+        return false;
+    }
+
+    return calibrate(plType, leftPupil.x, leftPupil.y, rightPupil.x, rightPupil.y, gaze.x, bcx, errorMsg)
+        && calibrate(plType, leftPupil.x, leftPupil.y, rightPupil.x, rightPupil.y, gaze.y, bcy, errorMsg);
+}
+
 bool BinocularPolyFit::calibrate(const PlType& type, const Mat& lx, const Mat& ly, const Mat& rx, const Mat& ry, const Mat& z, Mat1d& c, QString& errorMsg)
 {
 

EyeRecToo/src/gaze-estimation/BinocularPolyFit.h

@@ -18,6 +18,7 @@ private:
     cv::Mat1d bcy;
 
     bool calibrate(const PlType& type, const cv::Mat& lx, const cv::Mat& ly, const cv::Mat& rx, const cv::Mat& ry, const cv::Mat& z, cv::Mat1d& c, QString& errorMsg);
+    bool calibrateWithInliers(std::vector<CollectionTuple>& calibrationTuples, QString& errorMsg);
 };
 
 #endif // BINOCULARPOLYFIT_H