diff --git a/modules/video/doc/motion_analysis_and_object_tracking.rst b/modules/video/doc/motion_analysis_and_object_tracking.rst index 6c196c2ff..ebb9290cc 100644 --- a/modules/video/doc/motion_analysis_and_object_tracking.rst +++ b/modules/video/doc/motion_analysis_and_object_tracking.rst @@ -597,6 +597,48 @@ Returns background image See :ocv:func:`BackgroundSubtractor::getBackgroundImage`. +calcOpticalFlowSF +----------- +Calculate an optical flow using "SimpleFlow" algorithm. + +.. ocv:function:: void calcOpticalFlowSF( Mat& prev, Mat& next, Mat& flowX, Mat& flowY, int layers, int averaging_block_size, int max_flow, double sigma_dist, double sigma_color, int postprocess_window, double sigma_dist_fix, double sigma_color_fix, double occ_thr, int upscale_averaging_radiud, double upscale_sigma_dist, double upscale_sigma_color, double speed_up_thr) + + :param prev: First 8-bit 3-channel image. + + :param next: Second 8-bit 3-channel image + + :param flowX: X-coordinate of estimated flow + + :param flowY: Y-coordinate of estimated flow + + :param layers: Number of layers + + :param averaging_block_size: Size of block through which we sum up when calculate cost function for pixel + + :param max_flow: maximal flow that we search at each level + + :param sigma_dist: vector smooth spatial sigma parameter + + :param sigma_color: vector smooth color sigma parameter + + :param postprocess_window: window size for postprocess cross bilateral filter + + :param sigma_dist_fix: spatial sigma for postprocess cross bilateralf filter + + :param sigma_color_fix: color sigma for postprocess cross bilateral filter + + :param occ_thr: threshold for detecting occlusions + + :param upscale_averaging_radiud: window size for bilateral upscale operation + + :param upscale_sigma_dist: spatial sigma for bilateral upscale operation + + :param upscale_sigma_color: color sigma for bilateral upscale operation + + :param speed_up_thr: threshold to detect point with irregular flow - where flow should be recalculated after upscale + +See [Tao2012]_. And site of project - http://graphics.berkeley.edu/papers/Tao-SAN-2012-05/. + .. [Bouguet00] Jean-Yves Bouguet. Pyramidal Implementation of the Lucas Kanade Feature Tracker. .. [Bradski98] Bradski, G.R. "Computer Vision Face Tracking for Use in a Perceptual User Interface", Intel, 1998 @@ -612,3 +654,5 @@ See :ocv:func:`BackgroundSubtractor::getBackgroundImage`. .. [Lucas81] Lucas, B., and Kanade, T. An Iterative Image Registration Technique with an Application to Stereo Vision, Proc. of 7th International Joint Conference on Artificial Intelligence (IJCAI), pp. 674-679. .. [Welch95] Greg Welch and Gary Bishop “An Introduction to the Kalman Filter”, 1995 + +.. [Tao2012] Michael Tao, Jiamin Bai, Pushmeet Kohli and Sylvain Paris. SimpleFlow: A Non-iterative, Sublinear Optical Flow Algorithm. Computer Graphics Forum (Eurographics 2012) diff --git a/modules/video/include/opencv2/video/tracking.hpp b/modules/video/include/opencv2/video/tracking.hpp index 75668d228..85c18817a 100644 --- a/modules/video/include/opencv2/video/tracking.hpp +++ b/modules/video/include/opencv2/video/tracking.hpp @@ -326,7 +326,26 @@ CV_EXPORTS_W void calcOpticalFlowFarneback( InputArray prev, InputArray next, // that maps one 2D point set to another or one image to another. CV_EXPORTS_W Mat estimateRigidTransform( InputArray src, InputArray dst, bool fullAffine); - + +//! computes dense optical flow using Simple Flow algorithm +CV_EXPORTS_W void calcOpticalFlowSF(Mat& from, + Mat& to, + Mat& flowX, + Mat& flowY, + int layers, + int averaging_block_size, + int max_flow, + double sigma_dist, + double sigma_color, + int postprocess_window, + double sigma_dist_fix, + double sigma_color_fix, + double occ_thr, + int upscale_averaging_radius, + double upscale_sigma_dist, + double upscale_sigma_color, + double speed_up_thr); + } #endif diff --git a/modules/video/src/simpleflow.cpp b/modules/video/src/simpleflow.cpp new file mode 100644 index 000000000..1fda3618c --- /dev/null +++ b/modules/video/src/simpleflow.cpp @@ -0,0 +1,757 @@ +/*M/////////////////////////////////////////////////////////////////////////////////////// +// +// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING. +// +// By downloading, copying, installing or using the software you agree to this license. +// If you do not agree to this license, do not download, install, +// copy or use the software. +// +// +// License Agreement +// For Open Source Computer Vision Library +// +// Copyright (C) 2000-2008, Intel Corporation, all rights reserved. +// Copyright (C) 2009, Willow Garage Inc., all rights reserved. +// Third party copyrights are property of their respective owners. +// +// Redistribution and use in source and binary forms, with or without modification, +// are permitted provided that the following conditions are met: +// +// * Redistribution's of source code must retain the above copyright notice, +// this list of conditions and the following disclaimer. +// +// * Redistribution's in binary form must reproduce the above copyright notice, +// this list of conditions and the following disclaimer in the documentation +// and/or other materials provided with the distribution. +// +// * The name of the copyright holders may not be used to endorse or promote products +// derived from this software without specific prior written permission. +// +// This software is provided by the copyright holders and contributors "as is" and +// any express or implied warranties, including, but not limited to, the implied +// warranties of merchantability and fitness for a particular purpose are disclaimed. +// In no event shall the Intel Corporation or contributors be liable for any direct, +// indirect, incidental, special, exemplary, or consequential damages +// (including, but not limited to, procurement of substitute goods or services; +// loss of use, data, or profits; or business interruption) however caused +// and on any theory of liability, whether in contract, strict liability, +// or tort (including negligence or otherwise) arising in any way out of +// the use of this software, even if advised of the possibility of such damage. +// +//M*/ + +#include "precomp.hpp" +#include "simpleflow.hpp" + +// +// 2D dense optical flow algorithm from the following paper: +// Michael Tao, Jiamin Bai, Pushmeet Kohli, and Sylvain Paris. +// "SimpleFlow: A Non-iterative, Sublinear Optical Flow Algorithm" +// Computer Graphics Forum (Eurographics 2012) +// http://graphics.berkeley.edu/papers/Tao-SAN-2012-05/ +// + +namespace cv +{ + +WeightedCrossBilateralFilter::WeightedCrossBilateralFilter( + const Mat& _image, + int _windowSize, + double _sigmaDist, + double _sigmaColor) + : image(_image), + windowSize(_windowSize), + sigmaDist(_sigmaDist), + sigmaColor(_sigmaColor) { + + expDist.resize(2*windowSize*windowSize+1); + const double sigmaDistSqr = 2 * sigmaDist * sigmaDist; + for (int i = 0; i <= 2*windowSize*windowSize; ++i) { + expDist[i] = exp(-i/sigmaDistSqr); + } + + const double sigmaColorSqr = 2 * sigmaColor * sigmaColor; + wc.resize(image.rows); + for (int row = 0; row < image.rows; ++row) { + wc[row].resize(image.cols); + for (int col = 0; col < image.cols; ++col) { + int beginRow = max(0, row - windowSize); + int beginCol = max(0, col - windowSize); + int endRow = min(image.rows - 1, row + windowSize); + int endCol = min(image.cols - 1, col + windowSize); + wc[row][col] = build(endRow - beginRow + 1, endCol - beginCol + 1); + + for (int r = beginRow; r <= endRow; ++r) { + for (int c = beginCol; c <= endCol; ++c) { + wc[row][col][r - beginRow][c - beginCol] = + exp(-dist(image.at(row, col), + image.at(r, c)) + / sigmaColorSqr); + } + } + } + } +} + +Mat WeightedCrossBilateralFilter::apply(Mat& matrix, Mat& weights) { + int rows = matrix.rows; + int cols = matrix.cols; + + Mat result = Mat::zeros(rows, cols, CV_64F); + for (int row = 0; row < rows; ++row) { + for(int col = 0; col < cols; ++col) { + result.at(row, col) = + convolution(matrix, row, col, weights); + } + } + return result; +} + +double WeightedCrossBilateralFilter::convolution(Mat& matrix, + int row, int col, + Mat& weights) { + double result = 0, weightsSum = 0; + int beginRow = max(0, row - windowSize); + int beginCol = max(0, col - windowSize); + int endRow = min(matrix.rows - 1, row + windowSize); + int endCol = min(matrix.cols - 1, col + windowSize); + for (int r = beginRow; r <= endRow; ++r) { + double* ptr = matrix.ptr(r); + for (int c = beginCol; c <= endCol; ++c) { + const double w = expDist[dist(row, col, r, c)] * + wc[row][col][r - beginRow][c - beginCol] * + weights.at(r, c); + result += ptr[c] * w; + weightsSum += w; + } + } + return result / weightsSum; +} + +static void removeOcclusions(const Flow& flow, + const Flow& flow_inv, + double occ_thr, + Mat& confidence) { + const int rows = flow.u.rows; + const int cols = flow.v.cols; + int occlusions = 0; + for (int r = 0; r < rows; ++r) { + for (int c = 0; c < cols; ++c) { + if (dist(flow.u.at(r, c), flow.v.at(r, c), + -flow_inv.u.at(r, c), -flow_inv.v.at(r, c)) > occ_thr) { + confidence.at(r, c) = 0; + occlusions++; + } + } + } +} + +static Mat wd(int top_shift, int bottom_shift, int left_shift, int right_shift, double sigma) { + const double factor = 1.0 / (2.0 * sigma * sigma); + Mat d = Mat(top_shift + bottom_shift + 1, right_shift + left_shift + 1, CV_64F); + for (int dr = -top_shift, r = 0; dr <= bottom_shift; ++dr, ++r) { + for (int dc = -left_shift, c = 0; dc <= right_shift; ++dc, ++c) { + d.at(r, c) = -(dr*dr + dc*dc) * factor; + } + } + Mat ed; + exp(d, ed); + return ed; +} + +static Mat wc(const Mat& image, int r0, int c0, int top_shift, int bottom_shift, int left_shift, int right_shift, double sigma) { + const double factor = 1.0 / (2.0 * sigma * sigma); + Mat d = Mat(top_shift + bottom_shift + 1, right_shift + left_shift + 1, CV_64F); + for (int dr = r0-top_shift, r = 0; dr <= r0+bottom_shift; ++dr, ++r) { + for (int dc = c0-left_shift, c = 0; dc <= c0+right_shift; ++dc, ++c) { + d.at(r, c) = -dist(image.at(r0, c0), image.at(dr, dc)) * factor; + } + } + Mat ed; + exp(d, ed); + return ed; +} + +inline static void dist(const Mat& m1, const Mat& m2, Mat& result) { + const int rows = m1.rows; + const int cols = m1.cols; + for (int r = 0; r < rows; ++r) { + const Vec3b *m1_row = m1.ptr(r); + const Vec3b *m2_row = m2.ptr(r); + double* row = result.ptr(r); + for (int c = 0; c < cols; ++c) { + row[c] = dist(m1_row[c], m2_row[c]); + } + } +} + +static void calcOpticalFlowSingleScaleSF(const Mat& prev, + const Mat& next, + const Mat& mask, + Flow& flow, + Mat& confidence, + int averaging_radius, + int max_flow, + double sigma_dist, + double sigma_color) { + const int rows = prev.rows; + const int cols = prev.cols; + confidence = Mat::zeros(rows, cols, CV_64F); + + for (int r0 = 0; r0 < rows; ++r0) { + for (int c0 = 0; c0 < cols; ++c0) { + int u0 = floor(flow.u.at(r0, c0) + 0.5); + int v0 = floor(flow.v.at(r0, c0) + 0.5); + + const int min_row_shift = -min(r0 + u0, max_flow); + const int max_row_shift = min(rows - 1 - (r0 + u0), max_flow); + const int min_col_shift = -min(c0 + v0, max_flow); + const int max_col_shift = min(cols - 1 - (c0 + v0), max_flow); + + double min_cost = DBL_MAX, best_u = u0, best_v = v0; + + Mat w_full_window; + double w_full_window_sum; + Mat diff_storage; + + if (r0 - averaging_radius >= 0 && + r0 + averaging_radius < rows && + c0 - averaging_radius >= 0 && + c0 + averaging_radius < cols && + mask.at(r0, c0)) { + w_full_window = wd(averaging_radius, + averaging_radius, + averaging_radius, + averaging_radius, + sigma_dist).mul( + wc(prev, r0, c0, + averaging_radius, + averaging_radius, + averaging_radius, + averaging_radius, + sigma_color)); + + w_full_window_sum = sum(w_full_window)[0]; + diff_storage = Mat::zeros(averaging_radius*2 + 1, averaging_radius*2 + 1, CV_64F); + } + + bool first_flow_iteration = true; + double sum_e, min_e; + + for (int u = min_row_shift; u <= max_row_shift; ++u) { + for (int v = min_col_shift; v <= max_col_shift; ++v) { + double e = dist(prev.at(r0, c0), next.at(r0 + u0 + u, c0 + v0 + v)); + if (first_flow_iteration) { + sum_e = e; + min_e = e; + first_flow_iteration = false; + } else { + sum_e += e; + min_e = std::min(min_e, e); + } + if (!mask.at(r0, c0)) { + continue; + } + + const int window_top_shift = min(r0, r0 + u + u0, averaging_radius); + const int window_bottom_shift = min(rows - 1 - r0, + rows - 1 - (r0 + u + u0), + averaging_radius); + const int window_left_shift = min(c0, c0 + v + v0, averaging_radius); + const int window_right_shift = min(cols - 1 - c0, + cols - 1 - (c0 + v + v0), + averaging_radius); + + const Range prev_row_range(r0 - window_top_shift, r0 + window_bottom_shift + 1); + const Range prev_col_range(c0 - window_left_shift, c0 + window_right_shift + 1); + + const Range next_row_range(r0 + u0 + u - window_top_shift, + r0 + u0 + u + window_bottom_shift + 1); + const Range next_col_range(c0 + v0 + v - window_left_shift, + c0 + v0 + v + window_right_shift + 1); + + Mat diff2; + Mat w; + double w_sum; + if (window_top_shift == averaging_radius && + window_bottom_shift == averaging_radius && + window_left_shift == averaging_radius && + window_right_shift == averaging_radius) { + w = w_full_window; + w_sum = w_full_window_sum; + diff2 = diff_storage; + + dist(prev(prev_row_range, prev_col_range), next(next_row_range, next_col_range), diff2); + } else { + diff2 = Mat::zeros(window_bottom_shift + window_top_shift + 1, + window_right_shift + window_left_shift + 1, CV_64F); + + dist(prev(prev_row_range, prev_col_range), next(next_row_range, next_col_range), diff2); + + w = wd(window_top_shift, window_bottom_shift, window_left_shift, window_right_shift, sigma_dist).mul( + wc(prev, r0, c0, window_top_shift, window_bottom_shift, window_left_shift, window_right_shift, sigma_color)); + w_sum = sum(w)[0]; + } + multiply(diff2, w, diff2); + + const double cost = sum(diff2)[0] / w_sum; + if (cost < min_cost) { + min_cost = cost; + best_u = u + u0; + best_v = v + v0; + } + } + } + int square = (max_row_shift - min_row_shift + 1) * + (max_col_shift - min_col_shift + 1); + confidence.at(r0, c0) = (square == 0) ? 0 + : sum_e / square - min_e; + if (mask.at(r0, c0)) { + flow.u.at(r0, c0) = best_u; + flow.v.at(r0, c0) = best_v; + } + } + } +} + +static Flow upscaleOpticalFlow(int new_rows, + int new_cols, + const Mat& image, + const Mat& confidence, + const Flow& flow, + int averaging_radius, + double sigma_dist, + double sigma_color) { + const int rows = image.rows; + const int cols = image.cols; + Flow new_flow(new_rows, new_cols); + for (int r = 0; r < rows; ++r) { + for (int c = 0; c < cols; ++c) { + const int window_top_shift = min(r, averaging_radius); + const int window_bottom_shift = min(rows - 1 - r, averaging_radius); + const int window_left_shift = min(c, averaging_radius); + const int window_right_shift = min(cols - 1 - c, averaging_radius); + + const Range row_range(r - window_top_shift, r + window_bottom_shift + 1); + const Range col_range(c - window_left_shift, c + window_right_shift + 1); + + const Mat w = confidence(row_range, col_range).mul( + wd(window_top_shift, window_bottom_shift, window_left_shift, window_right_shift, sigma_dist)).mul( + wc(image, r, c, window_top_shift, window_bottom_shift, window_left_shift, window_right_shift, sigma_color)); + + const double w_sum = sum(w)[0]; + double new_u, new_v; + if (fabs(w_sum) < 1e-9) { + new_u = flow.u.at(r, c); + new_v = flow.v.at(r, c); + } else { + new_u = sum(flow.u(row_range, col_range).mul(w))[0] / w_sum; + new_v = sum(flow.v(row_range, col_range).mul(w))[0] / w_sum; + } + + for (int dr = 0; dr <= 1; ++dr) { + int nr = 2*r + dr; + for (int dc = 0; dc <= 1; ++dc) { + int nc = 2*c + dc; + if (nr < new_rows && nc < new_cols) { + new_flow.u.at(nr, nc) = 2 * new_u; + new_flow.v.at(nr, nc) = 2 * new_v; + } + } + } + } + } + return new_flow; +} + +static Mat calcIrregularityMat(const Flow& flow, int radius) { + const int rows = flow.u.rows; + const int cols = flow.v.cols; + Mat irregularity = Mat::zeros(rows, cols, CV_64F); + for (int r = 0; r < rows; ++r) { + const int start_row = max(0, r - radius); + const int end_row = min(rows - 1, r + radius); + for (int c = 0; c < cols; ++c) { + const int start_col = max(0, c - radius); + const int end_col = min(cols - 1, c + radius); + for (int dr = start_row; dr <= end_row; ++dr) { + for (int dc = start_col; dc <= end_col; ++dc) { + const double diff = dist(flow.u.at(r, c), flow.v.at(r, c), + flow.u.at(dr, dc), flow.v.at(dr, dc)); + if (diff > irregularity.at(r, c)) { + irregularity.at(r, c) = diff; + } + } + } + } + } + return irregularity; +} + +static void selectPointsToRecalcFlow(const Flow& flow, + int irregularity_metric_radius, + int speed_up_thr, + int curr_rows, + int curr_cols, + const Mat& prev_speed_up, + Mat& speed_up, + Mat& mask) { + const int prev_rows = flow.u.rows; + const int prev_cols = flow.v.cols; + + Mat is_flow_regular = calcIrregularityMat(flow, + irregularity_metric_radius) + < speed_up_thr; + Mat done = Mat::zeros(prev_rows, prev_cols, CV_8U); + speed_up = Mat::zeros(curr_rows, curr_cols, CV_8U); + mask = Mat::zeros(curr_rows, curr_cols, CV_8U); + + for (int r = 0; r < is_flow_regular.rows; ++r) { + for (int c = 0; c < is_flow_regular.cols; ++c) { + if (!done.at(r, c)) { + if (is_flow_regular.at(r, c) && + 2*r + 1 < curr_rows && 2*c + 1< curr_cols) { + + bool all_flow_in_region_regular = true; + int speed_up_at_this_point = prev_speed_up.at(r, c); + int step = (1 << speed_up_at_this_point) - 1; + int prev_top = r; + int prev_bottom = std::min(r + step, prev_rows - 1); + int prev_left = c; + int prev_right = std::min(c + step, prev_cols - 1); + + for (int rr = prev_top; rr <= prev_bottom; ++rr) { + for (int cc = prev_left; cc <= prev_right; ++cc) { + done.at(rr, cc) = 1; + if (!is_flow_regular.at(rr, cc)) { + all_flow_in_region_regular = false; + } + } + } + + int curr_top = std::min(2 * r, curr_rows - 1); + int curr_bottom = std::min(2*(r + step) + 1, curr_rows - 1); + int curr_left = std::min(2 * c, curr_cols - 1); + int curr_right = std::min(2*(c + step) + 1, curr_cols - 1); + + if (all_flow_in_region_regular && + curr_top != curr_bottom && + curr_left != curr_right) { + mask.at(curr_top, curr_left) = MASK_TRUE_VALUE; + mask.at(curr_bottom, curr_left) = MASK_TRUE_VALUE; + mask.at(curr_top, curr_right) = MASK_TRUE_VALUE; + mask.at(curr_bottom, curr_right) = MASK_TRUE_VALUE; + for (int rr = curr_top; rr <= curr_bottom; ++rr) { + for (int cc = curr_left; cc <= curr_right; ++cc) { + speed_up.at(rr, cc) = speed_up_at_this_point + 1; + } + } + } else { + for (int rr = curr_top; rr <= curr_bottom; ++rr) { + for (int cc = curr_left; cc <= curr_right; ++cc) { + mask.at(rr, cc) = MASK_TRUE_VALUE; + } + } + } + } else { + done.at(r, c) = 1; + for (int dr = 0; dr <= 1; ++dr) { + int nr = 2*r + dr; + for (int dc = 0; dc <= 1; ++dc) { + int nc = 2*c + dc; + if (nr < curr_rows && nc < curr_cols) { + mask.at(nr, nc) = MASK_TRUE_VALUE; + } + } + } + } + } + } + } +} + +static inline double extrapolateValueInRect(int height, int width, + double v11, double v12, + double v21, double v22, + int r, int c) { + if (r == 0 && c == 0) { return v11;} + if (r == 0 && c == width) { return v12;} + if (r == height && c == 0) { return v21;} + if (r == height && c == width) { return v22;} + + double qr = double(r) / height; + double pr = 1.0 - qr; + double qc = double(c) / width; + double pc = 1.0 - qc; + + return v11*pr*pc + v12*pr*qc + v21*qr*pc + v22*qc*qr; +} + +static void extrapolateFlow(Flow& flow, + const Mat& speed_up) { + const int rows = flow.u.rows; + const int cols = flow.u.cols; + Mat done = Mat::zeros(rows, cols, CV_8U); + for (int r = 0; r < rows; ++r) { + for (int c = 0; c < cols; ++c) { + if (!done.at(r, c) && speed_up.at(r, c) > 1) { + int step = (1 << speed_up.at(r, c)) - 1; + int top = r; + int bottom = std::min(r + step, rows - 1); + int left = c; + int right = std::min(c + step, cols - 1); + + int height = bottom - top; + int width = right - left; + for (int rr = top; rr <= bottom; ++rr) { + for (int cc = left; cc <= right; ++cc) { + done.at(rr, cc) = 1; + flow.u.at(rr, cc) = extrapolateValueInRect( + height, width, + flow.u.at(top, left), + flow.u.at(top, right), + flow.u.at(bottom, left), + flow.u.at(bottom, right), + rr-top, cc-left); + + flow.v.at(rr, cc) = extrapolateValueInRect( + height, width, + flow.v.at(top, left), + flow.v.at(top, right), + flow.v.at(bottom, left), + flow.v.at(bottom, right), + rr-top, cc-left); + } + } + } + } + } +} + +static void buildPyramidWithResizeMethod(Mat& src, + vector& pyramid, + int layers, + int interpolation_type) { + pyramid.push_back(src); + for (int i = 1; i <= layers; ++i) { + Mat prev = pyramid[i - 1]; + if (prev.rows <= 1 || prev.cols <= 1) { + break; + } + + Mat next; + resize(prev, next, Size((prev.cols + 1) / 2, (prev.rows + 1) / 2), 0, 0, interpolation_type); + pyramid.push_back(next); + } +} + +static Flow calcOpticalFlowSF(Mat& from, + Mat& to, + int layers, + int averaging_block_size, + int max_flow, + double sigma_dist, + double sigma_color, + int postprocess_window, + double sigma_dist_fix, + double sigma_color_fix, + double occ_thr, + int upscale_averaging_radius, + double upscale_sigma_dist, + double upscale_sigma_color, + double speed_up_thr) { + vector pyr_from_images; + vector pyr_to_images; + + buildPyramidWithResizeMethod(from, pyr_from_images, layers - 1, INTER_CUBIC); + buildPyramidWithResizeMethod(to, pyr_to_images, layers - 1, INTER_CUBIC); +// buildPyramid(from, pyr_from_images, layers - 1, BORDER_WRAP); +// buildPyramid(to, pyr_to_images, layers - 1, BORDER_WRAP); + + if ((int)pyr_from_images.size() != layers) { + exit(1); + } + + if ((int)pyr_to_images.size() != layers) { + exit(1); + } + + Mat first_from_image = pyr_from_images[layers - 1]; + Mat first_to_image = pyr_to_images[layers - 1]; + + Mat mask = Mat::ones(first_from_image.rows, first_from_image.cols, CV_8U); + Mat mask_inv = Mat::ones(first_from_image.rows, first_from_image.cols, CV_8U); + + Flow flow(first_from_image.rows, first_from_image.cols); + Flow flow_inv(first_to_image.rows, first_to_image.cols); + + Mat confidence; + Mat confidence_inv; + + calcOpticalFlowSingleScaleSF(first_from_image, + first_to_image, + mask, + flow, + confidence, + averaging_block_size, + max_flow, + sigma_dist, + sigma_color); + + calcOpticalFlowSingleScaleSF(first_to_image, + first_from_image, + mask_inv, + flow_inv, + confidence_inv, + averaging_block_size, + max_flow, + sigma_dist, + sigma_color); + + removeOcclusions(flow, + flow_inv, + occ_thr, + confidence); + + removeOcclusions(flow_inv, + flow, + occ_thr, + confidence_inv); + + Mat speed_up = Mat::zeros(first_from_image.rows, first_from_image.cols, CV_8U); + Mat speed_up_inv = Mat::zeros(first_from_image.rows, first_from_image.cols, CV_8U); + + for (int curr_layer = layers - 2; curr_layer >= 0; --curr_layer) { + const Mat curr_from = pyr_from_images[curr_layer]; + const Mat curr_to = pyr_to_images[curr_layer]; + const Mat prev_from = pyr_from_images[curr_layer + 1]; + const Mat prev_to = pyr_to_images[curr_layer + 1]; + + const int curr_rows = curr_from.rows; + const int curr_cols = curr_from.cols; + + Mat new_speed_up, new_speed_up_inv; + + selectPointsToRecalcFlow(flow, + averaging_block_size, + speed_up_thr, + curr_rows, + curr_cols, + speed_up, + new_speed_up, + mask); + + int points_to_recalculate = sum(mask)[0] / MASK_TRUE_VALUE; + + selectPointsToRecalcFlow(flow_inv, + averaging_block_size, + speed_up_thr, + curr_rows, + curr_cols, + speed_up_inv, + new_speed_up_inv, + mask_inv); + + points_to_recalculate = sum(mask_inv)[0] / MASK_TRUE_VALUE; + + speed_up = new_speed_up; + speed_up_inv = new_speed_up_inv; + + flow = upscaleOpticalFlow(curr_rows, + curr_cols, + prev_from, + confidence, + flow, + upscale_averaging_radius, + upscale_sigma_dist, + upscale_sigma_color); + + flow_inv = upscaleOpticalFlow(curr_rows, + curr_cols, + prev_to, + confidence_inv, + flow_inv, + upscale_averaging_radius, + upscale_sigma_dist, + upscale_sigma_color); + + calcOpticalFlowSingleScaleSF(curr_from, + curr_to, + mask, + flow, + confidence, + averaging_block_size, + max_flow, + sigma_dist, + sigma_color); + + calcOpticalFlowSingleScaleSF(curr_to, + curr_from, + mask_inv, + flow_inv, + confidence_inv, + averaging_block_size, + max_flow, + sigma_dist, + sigma_color); + + extrapolateFlow(flow, speed_up); + extrapolateFlow(flow_inv, speed_up_inv); + + removeOcclusions(flow, flow_inv, occ_thr, confidence); + removeOcclusions(flow_inv, flow, occ_thr, confidence_inv); + } + + WeightedCrossBilateralFilter filter_postprocess(pyr_from_images[0], + postprocess_window, + sigma_dist_fix, + sigma_color_fix); + + flow.u = filter_postprocess.apply(flow.u, confidence); + flow.v = filter_postprocess.apply(flow.v, confidence); + + Mat blured_u, blured_v; + GaussianBlur(flow.u, blured_u, Size(3, 3), 5); + GaussianBlur(flow.v, blured_v, Size(3, 3), 5); + + return Flow(blured_v, blured_u); +} + +void calcOpticalFlowSF(Mat& from, + Mat& to, + Mat& flowX, + Mat& flowY, + int layers, + int averaging_block_size, + int max_flow, + double sigma_dist, + double sigma_color, + int postprocess_window, + double sigma_dist_fix, + double sigma_color_fix, + double occ_thr, + int upscale_averaging_radius, + double upscale_sigma_dist, + double upscale_sigma_color, + double speed_up_thr) { + + Flow flow = calcOpticalFlowSF(from, to, + layers, + averaging_block_size, + max_flow, + sigma_dist, + sigma_color, + postprocess_window, + sigma_dist_fix, + sigma_color_fix, + occ_thr, + upscale_averaging_radius, + upscale_sigma_dist, + upscale_sigma_color, + speed_up_thr); + flowX = flow.u; + flowY = flow.v; +} + +} + diff --git a/modules/video/src/simpleflow.hpp b/modules/video/src/simpleflow.hpp new file mode 100644 index 000000000..55052fd05 --- /dev/null +++ b/modules/video/src/simpleflow.hpp @@ -0,0 +1,125 @@ +/*M/////////////////////////////////////////////////////////////////////////////////////// +// +// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING. +// +// By downloading, copying, installing or using the software you agree to this license. +// If you do not agree to this license, do not download, install, +// copy or use the software. +// +// +// License Agreement +// For Open Source Computer Vision Library +// +// Copyright (C) 2000-2008, Intel Corporation, all rights reserved. +// Copyright (C) 2009, Willow Garage Inc., all rights reserved. +// Third party copyrights are property of their respective owners. +// +// Redistribution and use in source and binary forms, with or without modification, +// are permitted provided that the following conditions are met: +// +// * Redistribution's of source code must retain the above copyright notice, +// this list of conditions and the following disclaimer. +// +// * Redistribution's in binary form must reproduce the above copyright notice, +// this list of conditions and the following disclaimer in the documentation +// and/or other materials provided with the distribution. +// +// * The name of the copyright holders may not be used to endorse or promote products +// derived from this software without specific prior written permission. +// +// This software is provided by the copyright holders and contributors "as is" and +// any express or implied warranties, including, but not limited to, the implied +// warranties of merchantability and fitness for a particular purpose are disclaimed. +// In no event shall the Intel Corporation or contributors be liable for any direct, +// indirect, incidental, special, exemplary, or consequential damages +// (including, but not limited to, procurement of substitute goods or services; +// loss of use, data, or profits; or business interruption) however caused +// and on any theory of liability, whether in contract, strict liability, +// or tort (including negligence or otherwise) arising in any way out of +// the use of this software, even if advised of the possibility of such damage. +// +//M*/ + +#ifndef __OPENCV_SIMPLEFLOW_H__ +#define __OPENCV_SIMPLEFLOW_H__ + +#include + +using namespace std; + +#define MASK_TRUE_VALUE 255 +#define UNKNOWN_FLOW_THRESH 1e9 + +namespace cv { + +struct Flow { + Mat u, v; + + Flow() {;} + + Flow(Mat& _u, Mat& _v) + : u(_u), v(_v) {;} + + Flow(int rows, int cols) { + u = Mat::zeros(rows, cols, CV_64F); + v = Mat::zeros(rows, cols, CV_64F); + } +}; + +inline static double dist(const Vec3b& p1, const Vec3b& p2) { + return (p1[0] - p2[0]) * (p1[0] - p2[0]) + + (p1[1] - p2[1]) * (p1[1] - p2[1]) + + (p1[2] - p2[2]) * (p1[2] - p2[2]); +} + +inline static double dist(const Point2f& p1, const Point2f& p2) { + return (p1.x - p2.x) * (p1.x - p2.x) + + (p1.y - p2.y) * (p1.y - p2.y); +} + +inline static double dist(double x1, double y1, double x2, double y2) { + return (x1 - x2) * (x1 - x2) + + (y1 - y2) * (y1 - y2); +} + +inline static int dist(int x1, int y1, int x2, int y2) { + return (x1 - x2) * (x1 - x2) + + (y1 - y2) * (y1 - y2); +} + +template +inline static T min(T t1, T t2, T t3) { + return (t1 <= t2 && t1 <= t3) ? t1 : min(t2, t3); +} + +template +vector > build(int n, int m) { + vector > res(n); + for (int i = 0; i < n; ++i) { + res[i].resize(m, 0); + } + return res; +} + +class WeightedCrossBilateralFilter { +public: + WeightedCrossBilateralFilter(const Mat& _image, + int _windowSize, + double _sigmaDist, + double _sigmaColor); + + Mat apply(Mat& matrix, Mat& weights); + +private: + double convolution(Mat& matrix, int row, int col, Mat& weights); + + Mat image; + int windowSize; + double sigmaDist, sigmaColor; + + vector expDist; + vector > > > wc; +}; +} + +#endif diff --git a/modules/video/test/test_simpleflow.cpp b/modules/video/test/test_simpleflow.cpp new file mode 100644 index 000000000..186ba8f56 --- /dev/null +++ b/modules/video/test/test_simpleflow.cpp @@ -0,0 +1,193 @@ +/*M/////////////////////////////////////////////////////////////////////////////////////// +// +// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING. +// +// By downloading, copying, installing or using the software you agree to this license. +// If you do not agree to this license, do not download, install, +// copy or use the software. +// +// +// Intel License Agreement +// For Open Source Computer Vision Library +// +// Copyright (C) 2000, Intel Corporation, all rights reserved. +// Third party copyrights are property of their respective owners. +// +// Redistribution and use in source and binary forms, with or without modification, +// are permitted provided that the following conditions are met: +// +// * Redistribution's of source code must retain the above copyright notice, +// this list of conditions and the following disclaimer. +// +// * Redistribution's in binary form must reproduce the above copyright notice, +// this list of conditions and the following disclaimer in the documentation +// and/or other materials provided with the distribution. +// +// * The name of Intel Corporation may not be used to endorse or promote products +// derived from this software without specific prior written permission. +// +// This software is provided by the copyright holders and contributors "as is" and +// any express or implied warranties, including, but not limited to, the implied +// warranties of merchantability and fitness for a particular purpose are disclaimed. +// In no event shall the Intel Corporation or contributors be liable for any direct, +// indirect, incidental, special, exemplary, or consequential damages +// (including, but not limited to, procurement of substitute goods or services; +// loss of use, data, or profits; or business interruption) however caused +// and on any theory of liability, whether in contract, strict liability, +// or tort (including negligence or otherwise) arising in any way out of +// the use of this software, even if advised of the possibility of such damage. +// +//M*/ + +#include "test_precomp.hpp" + +#include + +using namespace std; + +/* ///////////////////// simpleflow_test ///////////////////////// */ + +class CV_SimpleFlowTest : public cvtest::BaseTest +{ +public: + CV_SimpleFlowTest(); +protected: + void run(int); +}; + + +CV_SimpleFlowTest::CV_SimpleFlowTest() {} + +static void readOpticalFlowFromFile(FILE* file, cv::Mat& flowX, cv::Mat& flowY) { + char header[5]; + if (fread(header, 1, 4, file) < 4 && (string)header != "PIEH") { + return; + } + + int cols, rows; + if (fread(&cols, sizeof(int), 1, file) != 1|| + fread(&rows, sizeof(int), 1, file) != 1) { + return; + } + + flowX = cv::Mat::zeros(rows, cols, CV_64F); + flowY = cv::Mat::zeros(rows, cols, CV_64F); + + for (int i = 0; i < rows; ++i) { + for (int j = 0; j < cols; ++j) { + float uPoint, vPoint; + if (fread(&uPoint, sizeof(float), 1, file) != 1 || + fread(&vPoint, sizeof(float), 1, file) != 1) { + flowX.release(); + flowY.release(); + return; + } + + flowX.at(i, j) = uPoint; + flowY.at(i, j) = vPoint; + } + } +} + +static bool isFlowCorrect(double u) { + return !isnan(u) && (fabs(u) < 1e9); +} + +static double calc_rmse(cv::Mat flow1X, cv::Mat flow1Y, cv::Mat flow2X, cv::Mat flow2Y) { + long double sum; + int counter = 0; + const int rows = flow1X.rows; + const int cols = flow1X.cols; + + for (int y = 0; y < rows; ++y) { + for (int x = 0; x < cols; ++x) { + double u1 = flow1X.at(y, x); + double v1 = flow1Y.at(y, x); + double u2 = flow2X.at(y, x); + double v2 = flow2Y.at(y, x); + if (isFlowCorrect(u1) && isFlowCorrect(u2) && isFlowCorrect(v1) && isFlowCorrect(v2)) { + sum += (u1-u2)*(u1-u2) + (v1-v2)*(v1-v2); + counter++; + } + } + } + return sqrt((double)sum / (1e-9 + counter)); +} + +void CV_SimpleFlowTest::run(int) { + int code = cvtest::TS::OK; + + const double MAX_RMSE = 0.6; + const string frame1_path = ts->get_data_path() + "optflow/RubberWhale1.png"; + const string frame2_path = ts->get_data_path() + "optflow/RubberWhale2.png"; + const string gt_flow_path = ts->get_data_path() + "optflow/RubberWhale.flo"; + + cv::Mat frame1 = cv::imread(frame1_path); + cv::Mat frame2 = cv::imread(frame2_path); + + if (frame1.empty()) { + ts->printf(cvtest::TS::LOG, "could not read image %s\n", frame2_path.c_str()); + ts->set_failed_test_info(cvtest::TS::FAIL_MISSING_TEST_DATA); + return; + } + + if (frame2.empty()) { + ts->printf(cvtest::TS::LOG, "could not read image %s\n", frame2_path.c_str()); + ts->set_failed_test_info(cvtest::TS::FAIL_MISSING_TEST_DATA); + return; + } + + if (frame1.rows != frame2.rows && frame1.cols != frame2.cols) { + ts->printf(cvtest::TS::LOG, "images should be of equal sizes (%s and %s)", + frame1_path.c_str(), frame2_path.c_str()); + ts->set_failed_test_info(cvtest::TS::FAIL_MISSING_TEST_DATA); + return; + } + + if (frame1.type() != 16 || frame2.type() != 16) { + ts->printf(cvtest::TS::LOG, "images should be of equal type CV_8UC3 (%s and %s)", + frame1_path.c_str(), frame2_path.c_str()); + ts->set_failed_test_info(cvtest::TS::FAIL_MISSING_TEST_DATA); + return; + } + + cv::Mat flowX_gt, flowY_gt; + + FILE* gt_flow_file = fopen(gt_flow_path.c_str(), "rb"); + if (gt_flow_file == NULL) { + ts->printf(cvtest::TS::LOG, "could not read ground-thuth flow from file %s", + gt_flow_path.c_str()); + ts->set_failed_test_info(cvtest::TS::FAIL_MISSING_TEST_DATA); + return; + } + readOpticalFlowFromFile(gt_flow_file, flowX_gt, flowY_gt); + if (flowX_gt.empty() || flowY_gt.empty()) { + ts->printf(cvtest::TS::LOG, "error while reading flow data from file %s", + gt_flow_path.c_str()); + ts->set_failed_test_info(cvtest::TS::FAIL_MISSING_TEST_DATA); + return; + } + fclose(gt_flow_file); + + cv::Mat flowX, flowY; + cv::calcOpticalFlowSF(frame1, frame2, + flowX, flowY, + 3, 4, 2, 4.1, 25.5, 18, 55.0, 25.5, 0.35, 18, 55.0, 25.5, 10); + + double rmse = calc_rmse(flowX_gt, flowY_gt, flowX, flowY); + + ts->printf(cvtest::TS::LOG, "Optical flow estimation RMSE for SimpleFlow algorithm : %lf\n", + rmse); + + if (rmse > MAX_RMSE) { + ts->printf( cvtest::TS::LOG, + "Too big rmse error : %lf ( >= %lf )\n", rmse, MAX_RMSE); + ts->set_failed_test_info(cvtest::TS::FAIL_BAD_ACCURACY); + return; + } +} + + +TEST(Video_OpticalFlowSimpleFlow, accuracy) { CV_SimpleFlowTest test; test.safe_run(); } + +/* End of file. */ diff --git a/samples/cpp/simpleflow_demo.cpp b/samples/cpp/simpleflow_demo.cpp new file mode 100644 index 000000000..6a195fe89 --- /dev/null +++ b/samples/cpp/simpleflow_demo.cpp @@ -0,0 +1,96 @@ +#include "opencv2/video/tracking.hpp" +#include "opencv2/imgproc/imgproc.hpp" +#include "opencv2/highgui/highgui.hpp" + +#include +#include + +using namespace cv; +using namespace std; + +static void help() +{ + // print a welcome message, and the OpenCV version + printf("This is a demo of SimpleFlow optical flow algorithm,\n" + "Using OpenCV version %s\n\n", CV_VERSION); + + printf("Usage: simpleflow_demo frame1 frame2 output_flow" + "\nApplication will write estimated flow " + "\nbetween 'frame1' and 'frame2' in binary format" + "\ninto file 'output_flow'" + "\nThen one can use code from http://vision.middlebury.edu/flow/data/" + "\nto convert flow in binary file to image\n"); +} + +// binary file format for flow data specified here: +// http://vision.middlebury.edu/flow/data/ +static void writeOpticalFlowToFile(const Mat& u, const Mat& v, FILE* file) { + int cols = u.cols; + int rows = u.rows; + + fprintf(file, "PIEH"); + + if (fwrite(&cols, sizeof(int), 1, file) != 1 || + fwrite(&rows, sizeof(int), 1, file) != 1) { + fprintf(stderr, "writeOpticalFlowToFile : problem writing header\n"); + exit(1); + } + + for (int i= 0; i < u.rows; ++i) { + for (int j = 0; j < u.cols; ++j) { + float uPoint = u.at(i, j); + float vPoint = v.at(i, j); + + if (fwrite(&uPoint, sizeof(float), 1, file) != 1 || + fwrite(&vPoint, sizeof(float), 1, file) != 1) { + fprintf(stderr, "writeOpticalFlowToFile : problem writing data\n"); + exit(1); + } + } + } +} +int main(int argc, char** argv) { + help(); + + if (argc < 4) { + fprintf(stderr, "Wrong number of command line arguments : %d (expected %d)\n", argc, 4); + exit(1); + } + + Mat frame1 = imread(argv[1]); + Mat frame2 = imread(argv[2]); + + if (frame1.empty() || frame2.empty()) { + fprintf(stderr, "simpleflow_demo : Images cannot be read\n"); + exit(1); + } + + if (frame1.rows != frame2.rows && frame1.cols != frame2.cols) { + fprintf(stderr, "simpleflow_demo : Images should be of equal sizes\n"); + exit(1); + } + + if (frame1.type() != 16 || frame2.type() != 16) { + fprintf(stderr, "simpleflow_demo : Images should be of equal type CV_8UC3\n"); + exit(1); + } + + printf("simpleflow_demo : Read two images of size [rows = %d, cols = %d]\n", + frame1.rows, frame1.cols); + + Mat flowX, flowY; + + calcOpticalFlowSF(frame1, frame2, + flowX, flowY, + 3, 2, 4, 4.1, 25.5, 18, 55.0, 25.5, 0.35, 18, 55.0, 25.5, 10); + + FILE* file = fopen(argv[3], "wb"); + if (file == NULL) { + fprintf(stderr, "simpleflow_demo : Unable to open file '%s' for writing\n", argv[3]); + exit(1); + } + printf("simpleflow_demo : Writing to file\n"); + writeOpticalFlowToFile(flowX, flowY, file); + fclose(file); + return 0; +}