int xOffset(Mat y_LR) { int w = y_LR.size().width / 2; int h = y_LR.size().height; Mat left(y_LR, Rect(0, 0, w, h)); Mat right(y_LR, Rect(w, 0, w, h)); Mat disp(h, w, CV_16S); disparity(left, right, disp); // now compute the average disparity Mat mask = (disp > -200) & (disp < 2000); // but not where it's out of range // FIXME compute the median instead Scalar avg = cv::mean(disp, mask); return avg[0]; }
static bool read_labels(const string& path, vector<string>& filenames, vector< vector<Rect> >& labels) { string labels_path = path + "/gt.txt"; string filename, line; int x1, y1, x2, y2; char delim; ifstream ifs(labels_path.c_str()); if( !ifs.good() ) return false; while( getline(ifs, line) ) { stringstream stream(line); stream >> filename; filenames.push_back(path + "/" + filename); vector<Rect> filename_labels; while( stream >> x1 >> y1 >> x2 >> y2 >> delim ) { filename_labels.push_back(Rect(x1, y1, x2, y2)); } labels.push_back(filename_labels); filename_labels.clear(); } return true; }
TEST(hole_filling_test, rectangular_hole_on_repeated_texture_should_give_good_result) { Mat img = imread("test_images/brick_pavement.jpg"); convert_for_computation(img, 0.5f); // Add some hole Mat hole_mask = Mat::zeros(img.size(), CV_8U); hole_mask(Rect(72, 65, 5, 20)) = 255; int patch_size = 7; HoleFilling hf(img, hole_mask, patch_size); // Dump image with hole as black region. Mat img_with_hole_bgr; cvtColor(img, img_with_hole_bgr, CV_Lab2BGR); img_with_hole_bgr.setTo(Scalar(0,0,0), hole_mask); imwrite("brick_pavement_hole.exr", img_with_hole_bgr); // Dump reconstructed image Mat filled = hf.run(); cvtColor(filled, filled, CV_Lab2BGR); imwrite("brick_pavement_hole_filled.exr", filled); // The reconstructed image should be close to the original one, in this very simple case. Mat img_bgr; cvtColor(img, img_bgr, CV_Lab2BGR); double ssd = norm(img_bgr, filled, cv::NORM_L2SQR); EXPECT_LT(ssd, 0.2); }
void DeepPyramid::processFeatureMap(int filterIdx, const FeatureMap &map, vector<BoundingBox> &detectedObjects) const { Size mapSize = map.size(); Size filterSize = rootFilter[filterIdx]->getMapSize(); cout << "size: "<<map.size()<<endl; for (int width = 0; width < mapSize.width-filterSize.width; width+=stride) { for (int height = 0; height < mapSize.height-filterSize.height; height+=stride) { FeatureMap extractedMap; map.extractFeatureMap(Rect(Point(width, height), filterSize), extractedMap); if (rootFilter[filterIdx]->predict(extractedMap) == OBJECT) { BoundingBox box; box.norm5Box = Rect(Point(width, height), filterSize); box.confidence = std::fabs(rootFilter[filterIdx]->predict(extractedMap, true)); box.map = extractedMap; detectedObjects.push_back(box); } } } }
Rect DetectorTrainer::createRandomBounds() const { typedef std::uniform_int_distribution<int> uniform_int; int minWidth = featureParams.windowSizeInPixels().width; int maxWidth = std::min(imageSize.width, static_cast<int>(imageSize.height * aspectRatio)); int width = uniform_int{minWidth, maxWidth}(generator); int height = static_cast<int>(std::round(width * aspectRatioInv)); int x = uniform_int{0, imageSize.width - width}(generator); int y = uniform_int{0, imageSize.height - height}(generator); return Rect(x, y, width, height); }
void DeepPyramid::constructImagePyramid(const Mat& img, vector<Mat>& imgPyramid) const { Size imgSize(img.cols, img.rows); cout << "Create image pyramid..." << endl; for (int level = 0; level < levelCount; level++) { Mat imgAtLevel(net->inputLayerSize(), CV_8UC3, Scalar::all(0)); Mat resizedImg; Size resizedImgSize = embeddedImageSize(imgSize, level); resize(img, resizedImg, resizedImgSize); resizedImg.copyTo(imgAtLevel(Rect(Point(0, 0), resizedImgSize))); imgPyramid.push_back(imgAtLevel); } cout << "Status: Success!" << endl; }
void DeepPyramid::extractFeatureMap(const Mat &img, vector<Rect> &objects, Size size, vector<FeatureMap> &omaps, vector<FeatureMap>& nmaps) { Size imgSize(img.cols, img.rows); vector<FeatureMap> featureMaps; constructFeatureMapPyramid(img, featureMaps); for (int i = 0; i < levelCount; i++) { vector<Rect> objectsAtLevel; for (size_t obj = 0; obj < objects.size(); obj++) { objectsAtLevel.push_back(originalRect2Norm5(objects[obj], i, imgSize)); } Size mapSize = featureMaps[i].size(); for (int w = 0; w < mapSize.width-size.width; w+=stride) for (int h = 0; h < mapSize.height-size.height; h+=stride) { bool isNegative = true; bool isPositive = false; for (size_t obj = 0; obj < objects.size(); obj++) { if (IOU(Rect(Point(w, h), size), objectsAtLevel[obj]) > 0.3) isNegative = false; if (IOU(Rect(Point(w, h), size), objectsAtLevel[obj]) > 0.7) isPositive = true; } FeatureMap map; if (isNegative) { featureMaps[i].extractFeatureMap(Rect(Point(w, h), size), map); nmaps.push_back(map); } if (isPositive) { featureMaps[i].extractFeatureMap(Rect(Point(w, h), size), map); omaps.push_back(map); } } } }
void DeepPyramid::detect(const vector<FeatureMap>& maps, vector<BoundingBox>& detectedObjects) const { for (size_t i = 0; i < rootFilter.size(); i++) for (size_t j = 0; j < levelCount; j++) { vector<BoundingBox> detectedObjectsAtLevel; Size size = maps[j].size(); double scale = 1 / pow(2.0, (levelCount - j -1)/2.0); size.width = size.width * scale; size.height = size.height * scale; FeatureMap map; maps[j].extractFeatureMap(Rect(Point(0, 0), size), map); processFeatureMap(i, map, detectedObjectsAtLevel); for (size_t k = 0; k < detectedObjectsAtLevel.size(); k++) { detectedObjectsAtLevel[k].level = j; detectedObjects.push_back(detectedObjectsAtLevel[k]); } } }
TEST(hole_filling_test, one_pixel_hole_on_random_image_should_produce_correct_target_rect) { // Make some random image data. Mat img = Mat(100, 100, CV_32FC1); randu(img, 0.f, 1.f); // Put a hole in middle. Mat hole = Mat::zeros(100, 100, CV_8U); // Set one pixel as hole hole(Rect(50, 50, 1, 1)) = 1; int patch_size = 7; HoleFilling hf(img, hole, patch_size); Rect expected_target_rect(Point(50 - 6, 50 - 6), Point(50 + 7, 50 + 7)); ASSERT_EQ(expected_target_rect, hf._target_rect_pyr[0]); }
void EnhancedStereo::computeDynamicProgramming() { // cout << "left" << endl; // left tableau init for (int v = 0; v < smallHeight(); v++) { int * tableauRow = (int *)(tableauLeft.row(v).data); uint8_t * errorRow = errorBuffer.row(v).data; // init the first row copy(errorRow, errorRow + dispMax, tableauRow); // fill up the tableau for (int u = 1; u < smallWidth(); u++) { computeDynamicStep(tableauRow + (u - 1)*dispMax, errorRow + u*dispMax, tableauRow + u*dispMax); } } // cout << "right" << endl; // right tableau init for (int v = 0; v < smallHeight(); v++) { int * tableauRow = (int *)(tableauRight.row(v).data); uint8_t * errorRow = errorBuffer.row(v).data; int base = (smallWidth() - 1) * dispMax; copy(errorRow + base, errorRow + base + dispMax, tableauRow + base); for (int u = smallWidth() - 2; u >= 0; u--) { computeDynamicStep(tableauRow + (u + 1)*dispMax, errorRow + u*dispMax, tableauRow + u*dispMax); } } // cout << "top" << endl; // top-down tableau init for (int u = 0; u < smallWidth(); u++) { auto tableauCol = tableauTop(Rect(u*dispMax, 0, dispMax, smallHeight())); auto errorCol = errorBuffer(Rect(u*dispMax, 0, dispMax, smallHeight())); copy(errorCol.data, errorCol.data + dispMax, (int*)(tableauCol.data)); for (int v = 1; v < smallHeight(); v++) { computeDynamicStep((int*)(tableauCol.row(v-1).data), errorCol.row(v).data, (int*)(tableauCol.row(v).data)); } } // cout << "bottom" << endl; // bottom-up tableau init for (int u = 0; u < smallWidth(); u++) { auto tableauCol = tableauBottom(Rect(u*dispMax, 0, dispMax, smallHeight())); auto errorCol = errorBuffer(Rect(u*dispMax, 0, dispMax, smallHeight())); int vLast = smallHeight() - 1; copy(errorCol.row(vLast).data, errorCol.row(vLast).data + dispMax, (int*)(tableauCol.row(vLast).data)); for (int v = smallHeight() - 2; v >= 0; v--) { computeDynamicStep((int*)(tableauCol.row(v+1).data), errorCol.row(v).data, (int*)(tableauCol.row(v).data)); } } }
void FingerTracker::Setup() { VideoCapture capture(0); if (!capture.isOpened()) { throw std::runtime_error("Could not start camera capture"); } int windowSize = 25; int Xpos = 200; int Ypos = 50; int update = 0; int buttonClicked = 0; namedWindow("RGB", CV_WINDOW_AUTOSIZE); createTrackbar("X", "RGB", &Xpos, 320, TrackBarCallback, (void*)&update); createTrackbar("Y", "RGB", &Ypos, 240, TrackBarCallback, (void*)&update); createTrackbar("Size", "RGB", &windowSize, 100, TrackBarCallback, (void*)&update); setMouseCallback("RGB", MouseCallback, (void*)&buttonClicked); Mat fingerWindowBackground, fingerWindowBackgroundGray; m_calibrationData.reset(new CalibrationData()); bool ticking = false; std::chrono::system_clock::time_point start = std::chrono::system_clock::now(); while (true) { Mat frame, frameHSV; if (capture.read(frame)) { flip(frame, frame, 1); pyrDown(frame, frame, Size(frame.cols / 2, frame.rows / 2)); Rect fingerWindow(Point(Xpos, Ypos), Size(windowSize, windowSize*3)); if (Xpos + windowSize >= frame.cols || Ypos + windowSize*3 >= frame.rows) { windowSize = 20; Xpos = 200; Ypos = 50; update = 0; } else if (buttonClicked == 1) { frame(fingerWindow).copyTo(fingerWindowBackground); cvtColor(fingerWindowBackground, fingerWindowBackgroundGray, CV_BGR2GRAY); buttonClicked = 0; update = 0; cvDestroyAllWindows(); } if (fingerWindowBackgroundGray.rows && !m_calibrationData->m_ready) { Mat diff, thd; absdiff(frame(fingerWindow), fingerWindowBackground, diff); std::vector<Mat> ch; split(diff, ch); threshold(ch[0], ch[0], m_calibrationDiffThreshold, 255, 0); threshold(ch[1], ch[1], m_calibrationDiffThreshold, 255, 0); threshold(ch[2], ch[2], m_calibrationDiffThreshold, 255, 0); thd = ch[0]; add(thd, ch[1], thd); add(thd, ch[2], thd); medianBlur(thd, thd, 5); Mat top, middle, bottom; Rect r1 = Rect(0, 0, thd.cols, thd.rows/3); Rect r2 = Rect(0, thd.rows / 3 + 1, thd.cols, thd.rows/3); Rect r3 = Rect(0, thd.rows * 2 / 3 + 1, thd.cols, thd.rows - thd.rows * 2 / 3 - 1); top = thd(r1); middle = thd(r2); bottom = thd(r3); auto percentageTop = countNonZero(top) * 100.0 / top.size().area(); auto percentageMiddle = countNonZero(middle) * 100.0 / middle.size().area(); auto percentageBottom = countNonZero(bottom) * 100.0 / bottom.size().area(); bool topReady = false; bool middleReady = false; bool bottomReady = false; Scalar c1, c2, c3; if (percentageTop > m_calibrationTopLowerThd && percentageTop < m_calibrationTopUpperThd) { topReady = true; c1 = Scalar(0, 255, 255); } else { c1 = Scalar(0, 0, 255); } if (percentageMiddle > m_calibrationMiddleLowerThd && percentageMiddle < m_calibrationMiddleUppperThd) { middleReady = true; c2 = Scalar(0, 255, 255); } else { c2 = Scalar(0, 0, 255); } if (percentageBottom > m_calibrationBottomLowerThd && percentageBottom < m_calibrationBottomUpperThd) { bottomReady = true; c3 = Scalar(0, 255, 255); } else { c3 = Scalar(0, 0, 255); } bool readyToGo = false; if (middleReady && topReady && bottomReady) { c1 = Scalar(0, 255, 0); c2 = Scalar(0, 255, 0); c3 = Scalar(0, 255, 0); if (!ticking) { start = std::chrono::system_clock::now(); ticking = true; } if (std::chrono::system_clock::now() - start > std::chrono::seconds(1)) { readyToGo = true; } } else { ticking = false; } #if ENABLE_DEBUG_WINDOWS std::stringstream ss; ss << percentageTop << ", " << percentageMiddle << ", " << percentageBottom; putText(frame, ss.str(), Point(0, getTextSize(ss.str(), 0, 0.5, 1, nullptr).height), 0, 0.5, Scalar::all(255), 1); cv::imshow("Thresholded", thd); #endif if (percentageTop >= m_calibrationTopUpperThd && percentageBottom >= m_calibrationBottomUpperThd && percentageMiddle >= m_calibrationMiddleUppperThd) { putText(frame, "Move finger away from camera", Point(0, getTextSize("Move finger away from camera", 0, 0.5, 1, nullptr).height), 0, 0.5, Scalar::all(255), 1); } else if (percentageTop <= m_calibrationTopLowerThd && percentageBottom <= m_calibrationBottomLowerThd && percentageMiddle <= m_calibrationMiddleLowerThd) { putText(frame, "Move finger closer to camera", Point(0, getTextSize("Move finger closer to camera", 0, 0.5, 1, nullptr).height), 0, 0.5, Scalar::all(255), 1); } if (readyToGo) { Mat framePatchHSV; cvtColor(frame(fingerWindow), framePatchHSV, CV_BGR2HSV); cvtColor(frame, frameHSV, CV_BGR2HSV); MatND hist; calcHist(&framePatchHSV, 1, m_calibrationData->m_channels, thd, hist, 2, m_calibrationData->m_histSize, (const float**)m_calibrationData->m_ranges, true, false); m_calibrationData->m_hist = hist; normalize(m_calibrationData->m_hist, m_calibrationData->m_hist, 0, 255, NORM_MINMAX, -1, Mat()); #if ENABLE_DEBUG_WINDOWS double maxVal=0; minMaxLoc(m_calibrationData->m_hist, 0, &maxVal, 0, 0); int scale = 10; Mat histImg = Mat::zeros(m_calibrationData->m_sbins*scale, m_calibrationData->m_hbins*10, CV_8UC3); for( int h = 0; h < m_calibrationData->m_hbins; h++) { for( int s = 0; s < m_calibrationData->m_sbins; s++ ) { float binVal = m_calibrationData->m_hist.at<float>(h, s); int intensity = cvRound(binVal*255/maxVal); rectangle( histImg, Point(h*scale, s*scale), Point( (h+1)*scale - 1, (s+1)*scale - 1), Scalar::all(intensity), CV_FILLED ); } } imshow("H-S Histogram", histImg); #endif m_calibrationData->m_ready = true; frame(fingerWindow).copyTo(m_calibrationData->m_fingerPatch); m_calibrationData->m_fingerRect = fingerWindow; m_currentCandidate.m_windowRect = fingerWindow; m_currentCandidate.m_fingerPosition = fingerWindow.tl(); return; } rectangle(frame, r1.tl() + fingerWindow.tl(), r1.br() + fingerWindow.tl(), c1); rectangle(frame, r2.tl() + fingerWindow.tl(), r2.br() + fingerWindow.tl(), c2); rectangle(frame, r3.tl() + fingerWindow.tl(), r3.br() + fingerWindow.tl(), c3); imshow("Calibration", frame); } else { int baseline = 0; putText(frame, "Adjust calibration window, click when ready", Point(0, getTextSize("Adjust calibration window", 0, 0.4, 2, &baseline).height), 0, 0.4, Scalar::all(255), 1); rectangle(frame, fingerWindow.tl(), fingerWindow.br(), Scalar(0, 0, 255)); imshow("RGB", frame); } auto key = cvWaitKey(10); if (char(key) == 27) { break; } } } capture.release(); }
void FingerTracker::Process(Mat frame) { #if ENABLE_DEBUG_WINDOWS Mat img_display; frame.copyTo(img_display); #endif // Process only Region of Interest i.e. region around current finger position Rect roi = Rect( Point(std::max(m_currentCandidate.m_windowRect.tl().x - m_roiSpanX, 0), std::max(m_currentCandidate.m_windowRect.tl().y - m_roiSpanY, 0)), Point(std::min(m_currentCandidate.m_windowRect.tl().x + m_roiSpanX + m_calibrationData->m_fingerPatch.cols, frame.cols), std::min(m_currentCandidate.m_windowRect.tl().y + m_roiSpanY + m_calibrationData->m_fingerPatch.rows, frame.rows))); Mat frameRoi; frame(roi).copyTo(frameRoi); //================TEMPLATE MATCHING int result_cols = frameRoi.cols - m_calibrationData->m_fingerPatch.cols + 1; int result_rows = frameRoi.rows - m_calibrationData->m_fingerPatch.rows + 1; assert(result_cols > 0 && result_rows > 0); Mat scoreMap; scoreMap.create(result_cols, result_rows, CV_32FC1); // Compare current frame roi region to known candidate // Using OpenCV matchTemplate function with correlation coefficient matching method matchTemplate(frameRoi, m_calibrationData->m_fingerPatch, scoreMap, 3); //================HISTOGRAM BACK PROJECTION MatND backProjection; Mat frameHSV; cvtColor(frameRoi, frameHSV, CV_BGR2HSV); calcBackProject(&frameHSV, 1, m_calibrationData->m_channels, m_calibrationData->m_hist, backProjection, (const float**)(m_calibrationData->m_ranges), 1, true); Mat backProjectionThresholded; threshold(backProjection, backProjectionThresholded, m_backProjectionThreshold, 255, 0); erode(backProjectionThresholded, backProjectionThresholded, getStructuringElement(MORPH_RECT, Size(2 * m_erosionSize + 1, 2 * m_erosionSize + 1), Point(m_erosionSize, m_erosionSize))); dilate(backProjectionThresholded, backProjectionThresholded, getStructuringElement(MORPH_RECT, Size(2 * m_dilationSize + 1, 2 * m_dilationSize + 1), Point(m_dilationSize, m_dilationSize))); Mat backProjectionThresholdedShifted; Rect shifted(Rect(m_calibrationData->m_fingerPatch.cols - 1, m_calibrationData->m_fingerPatch.rows - 1, scoreMap.cols, scoreMap.rows)); backProjectionThresholded(shifted).copyTo(backProjectionThresholdedShifted); Mat maskedOutScoreMap(scoreMap.size(), CV_8U); scoreMap.copyTo(maskedOutScoreMap, backProjectionThresholdedShifted); //====================Localizing the best match with minMaxLoc double minVal; double maxVal; Point minLoc; Point maxLoc; Point matchLoc; minMaxLoc(maskedOutScoreMap, &minVal, &maxVal, &minLoc, &maxLoc, Mat()); matchLoc = maxLoc + roi.tl(); m_currentCandidate.m_confidence = static_cast<float>(maxVal); if (maxVal > m_candidateDetecionConfidenceThreshold) { m_currentCandidate.m_found = true; m_currentCandidate.m_windowRect = Rect(matchLoc, Point(matchLoc.x + m_calibrationData->m_fingerPatch.cols , matchLoc.y + m_calibrationData->m_fingerPatch.rows)); //================Find finger position Mat fingerWindowThresholded; backProjectionThresholded(Rect(maxLoc, Point(maxLoc.x + m_calibrationData->m_fingerPatch.cols , maxLoc.y + m_calibrationData->m_fingerPatch.rows))).copyTo(fingerWindowThresholded); m_currentCandidate.m_fingerPosition = GetFingerTopPosition(fingerWindowThresholded) + matchLoc; #if ENABLE_DEBUG_WINDOWS rectangle(img_display, m_currentCandidate.m_windowRect.tl(), m_currentCandidate.m_windowRect.br(), Scalar(255,0,0), 2, 8, 0 ); rectangle(scoreMap, m_currentCandidate.m_windowRect.tl(), m_currentCandidate.m_windowRect.br(), Scalar::all(0), 2, 8, 0 ); rectangle(img_display, m_currentCandidate.m_fingerPosition, m_currentCandidate.m_fingerPosition + Point(5,5), Scalar(255,0,0)); #endif } else { m_currentCandidate.m_found = false; } #if ENABLE_DEBUG_WINDOWS std::stringstream ss; ss << maxVal; putText(img_display, ss.str(), Point(50, 50), 0, 0.5, Scalar(255,255,255), 1); imshow("Overlays", img_display); imshow("Results", scoreMap); imshow("ResultMasked", maskedOutScoreMap); #endif }
void getDistortionValues(cv::RNG &rng, const Size2i &inputSize, AugParams *agp) { // This function just gets the random distortion values without modifying the // image itself. Useful if we need to reapply the same transformations over // again (e.g. for all frames of a video or for a corresponding target mask) // colornoise values // N.B. if _contrastMax == 100, then _colorNoiseStd will be 0.0 for (int i=0; i<3; i++) { agp->colornoise[i] = rng.gaussian(_colorNoiseStd); } // contrast, brightness, saturation // N.B. all value ranges tied to _contrastMin and _contrastMax for (int i=0; i<3; i++) { agp->cbs[i] = rng.uniform(_contrastMin, _contrastMax) / 100.0f; } /************************** * HORIZONTAL FLIP * ***************************/ agp->flip = _flip && rng(2) != 0 ? true : false; /************************** * ROTATION ANGLE * ***************************/ agp->angle = rng.uniform(_rotateMin, _rotateMax); /************************** * CROP BOX * ***************************/ float shortSide = std::min(inputSize.height, inputSize.width); // Special case where we just grab the whole image; if (_scaleMin == 0) { agp->cropBox = Rect(Point2i(), inputSize); return; } if (_center) { agp->cropBox.width = shortSide * _width / (float) _scaleMin; agp->cropBox.height = shortSide * _height / (float) _scaleMin; agp->cropBox.x = (inputSize.width - agp->cropBox.width) / 2; agp->cropBox.y = (inputSize.height - agp->cropBox.height) / 2; } else { cv::Size2f oSize = inputSize; // This is a hack for backward compatibility. // Valid aspect ratio range ( > 100) will override side scaling behavior if (_aspectRatio == 0) { float scaleFactor = rng.uniform(_scaleMin, _scaleMax); agp->cropBox.width = shortSide * _width / scaleFactor; agp->cropBox.height = shortSide * _height / scaleFactor; } else { float mAR = (float) _aspectRatio / 100.0f; float nAR = rng.uniform(1.0f / mAR, mAR); float oAR = oSize.width / oSize.height; // between minscale pct% to 100% subject to aspect ratio limitation float maxScale = nAR > oAR ? oAR / nAR : nAR / oAR; float minScale = std::min((float) _scaleMin / 100.0f, maxScale); float tgtArea = rng.uniform(minScale, maxScale) * oSize.area(); agp->cropBox.height = sqrt(tgtArea / nAR); agp->cropBox.width = agp->cropBox.height * nAR; } agp->cropBox.x = rng.uniform(0, inputSize.width - agp->cropBox.width); agp->cropBox.y = rng.uniform(0, inputSize.height - agp->cropBox.height); } return; }
Mat get_L(Mat LR) { int w = LR.size().width / 2; int h = LR.size().height; return Mat(LR, Rect(0, 0, w, h)); }