From 3b0fa68a973113f398c217ca8402e82cc3e8867f Mon Sep 17 00:00:00 2001 From: vbystricky Date: Thu, 9 Jan 2014 16:34:01 +0400 Subject: [PATCH] Move OpticalFlowFarneback from ocl module to video module --- modules/video/src/opencl/optical_flow_farneback.cl | 434 +++++++++++++++++ modules/video/src/optflowgf.cpp | 526 ++++++++++++++++++++- modules/video/src/precomp.hpp | 1 + modules/video/test/ocl/test_optflow_farneback.cpp | 110 +++++ 4 files changed, 1053 insertions(+), 18 deletions(-) create mode 100644 modules/video/src/opencl/optical_flow_farneback.cl create mode 100644 modules/video/test/ocl/test_optflow_farneback.cpp diff --git a/modules/video/src/opencl/optical_flow_farneback.cl b/modules/video/src/opencl/optical_flow_farneback.cl new file mode 100644 index 0000000..0ef48d2 --- /dev/null +++ b/modules/video/src/opencl/optical_flow_farneback.cl @@ -0,0 +1,434 @@ +/*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) 2010-2012, Multicoreware, Inc., all rights reserved. +// Copyright (C) 2010-2012, Advanced Micro Devices, Inc., all rights reserved. +// Third party copyrights are property of their respective owners. +// +// @Authors +// Sen Liu, swjtuls1987@126.com +// +// 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*/ + + +#define tx (int)get_local_id(0) +#define ty get_local_id(1) +#define bx get_group_id(0) +#define bdx (int)get_local_size(0) + +#define BORDER_SIZE 5 +#define MAX_KSIZE_HALF 100 + +#ifndef polyN +#define polyN 5 +#endif + +#if USE_DOUBLE +#ifdef cl_amd_fp64 +#pragma OPENCL EXTENSION cl_amd_fp64:enable +#elif defined (cl_khr_fp64) +#pragma OPENCL EXTENSION cl_khr_fp64:enable +#endif +#define TYPE double +#define VECTYPE double4 +#else +#define TYPE float +#define VECTYPE float4 +#endif + +__kernel void polynomialExpansion(__global __const float * src, int srcStep, + __global float * dst, int dstStep, + const int rows, const int cols, + __global __const float * c_g, + __global __const float * c_xg, + __global __const float * c_xxg, + __local float * smem, + const VECTYPE ig) +{ + const int y = get_global_id(1); + const int x = bx * (bdx - 2*polyN) + tx - polyN; + + int xWarped; + __local float *row = smem + tx; + + if (y < rows && y >= 0) + { + xWarped = min(max(x, 0), cols - 1); + + row[0] = src[mad24(y, srcStep, xWarped)] * c_g[0]; + row[bdx] = 0.f; + row[2*bdx] = 0.f; + +#pragma unroll + for (int k = 1; k <= polyN; ++k) + { + float t0 = src[mad24(max(y - k, 0), srcStep, xWarped)]; + float t1 = src[mad24(min(y + k, rows - 1), srcStep, xWarped)]; + + row[0] += c_g[k] * (t0 + t1); + row[bdx] += c_xg[k] * (t1 - t0); + row[2*bdx] += c_xxg[k] * (t0 + t1); + } + } + + barrier(CLK_LOCAL_MEM_FENCE); + + if (y < rows && y >= 0 && tx >= polyN && tx + polyN < bdx && x < cols) + { + TYPE b1 = c_g[0] * row[0]; + TYPE b3 = c_g[0] * row[bdx]; + TYPE b5 = c_g[0] * row[2*bdx]; + TYPE b2 = 0, b4 = 0, b6 = 0; + +#pragma unroll + for (int k = 1; k <= polyN; ++k) + { + b1 += (row[k] + row[-k]) * c_g[k]; + b4 += (row[k] + row[-k]) * c_xxg[k]; + b2 += (row[k] - row[-k]) * c_xg[k]; + b3 += (row[k + bdx] + row[-k + bdx]) * c_g[k]; + b6 += (row[k + bdx] - row[-k + bdx]) * c_xg[k]; + b5 += (row[k + 2*bdx] + row[-k + 2*bdx]) * c_g[k]; + } + + dst[mad24(y, dstStep, xWarped)] = (float)(b3*ig.s0); + dst[mad24(rows + y, dstStep, xWarped)] = (float)(b2*ig.s0); + dst[mad24(2*rows + y, dstStep, xWarped)] = (float)(b1*ig.s1 + b5*ig.s2); + dst[mad24(3*rows + y, dstStep, xWarped)] = (float)(b1*ig.s1 + b4*ig.s2); + dst[mad24(4*rows + y, dstStep, xWarped)] = (float)(b6*ig.s3); + } +} + +inline int idx_row_low(const int y, const int last_row) +{ + return abs(y) % (last_row + 1); +} + +inline int idx_row_high(const int y, const int last_row) +{ + return abs(last_row - abs(last_row - y)) % (last_row + 1); +} + +inline int idx_row(const int y, const int last_row) +{ + return idx_row_low(idx_row_high(y, last_row), last_row); +} + +inline int idx_col_low(const int x, const int last_col) +{ + return abs(x) % (last_col + 1); +} + +inline int idx_col_high(const int x, const int last_col) +{ + return abs(last_col - abs(last_col - x)) % (last_col + 1); +} + +inline int idx_col(const int x, const int last_col) +{ + return idx_col_low(idx_col_high(x, last_col), last_col); +} + +__kernel void gaussianBlur(__global const float * src, int srcStep, + __global float * dst, int dstStep, const int rows, const int cols, + __global const float * c_gKer, const int ksizeHalf, + __local float * smem) +{ + const int y = get_global_id(1); + const int x = get_global_id(0); + + __local float *row = smem + ty * (bdx + 2*ksizeHalf); + + if (y < rows) + { + // Vertical pass + for (int i = tx; i < bdx + 2*ksizeHalf; i += bdx) + { + int xExt = (int)(bx * bdx) + i - ksizeHalf; + xExt = idx_col(xExt, cols - 1); + row[i] = src[mad24(y, srcStep, xExt)] * c_gKer[0]; + for (int j = 1; j <= ksizeHalf; ++j) + row[i] += (src[mad24(idx_row_low(y - j, rows - 1), srcStep, xExt)] + + src[mad24(idx_row_high(y + j, rows - 1), srcStep, xExt)]) * c_gKer[j]; + } + } + + barrier(CLK_LOCAL_MEM_FENCE); + + if (y < rows && y >= 0 && x < cols && x >= 0) + { + // Horizontal pass + row += tx + ksizeHalf; + float res = row[0] * c_gKer[0]; + for (int i = 1; i <= ksizeHalf; ++i) + res += (row[-i] + row[i]) * c_gKer[i]; + + dst[mad24(y, dstStep, x)] = res; + } +} + +__kernel void gaussianBlur5(__global const float * src, int srcStep, + __global float * dst, int dstStep, + const int rows, const int cols, + __global const float * c_gKer, const int ksizeHalf, + __local float * smem) +{ + const int y = get_global_id(1); + const int x = get_global_id(0); + + const int smw = bdx + 2*ksizeHalf; // shared memory "cols" + __local volatile float *row = smem + 5 * ty * smw; + + if (y < rows) + { + // Vertical pass + for (int i = tx; i < bdx + 2*ksizeHalf; i += bdx) + { + int xExt = (int)(bx * bdx) + i - ksizeHalf; + xExt = idx_col(xExt, cols - 1); + +#pragma unroll + for (int k = 0; k < 5; ++k) + row[k*smw + i] = src[mad24(k*rows + y, srcStep, xExt)] * c_gKer[0]; + + for (int j = 1; j <= ksizeHalf; ++j) +#pragma unroll + for (int k = 0; k < 5; ++k) + row[k*smw + i] += + (src[mad24(k*rows + idx_row_low(y - j, rows - 1), srcStep, xExt)] + + src[mad24(k*rows + idx_row_high(y + j, rows - 1), srcStep, xExt)]) * c_gKer[j]; + } + } + + barrier(CLK_LOCAL_MEM_FENCE); + + if (y < rows && y >= 0 && x < cols && x >= 0) + { + // Horizontal pass + + row += tx + ksizeHalf; + float res[5]; + +#pragma unroll + for (int k = 0; k < 5; ++k) + res[k] = row[k*smw] * c_gKer[0]; + + for (int i = 1; i <= ksizeHalf; ++i) +#pragma unroll + for (int k = 0; k < 5; ++k) + res[k] += (row[k*smw - i] + row[k*smw + i]) * c_gKer[i]; + +#pragma unroll + for (int k = 0; k < 5; ++k) + dst[mad24(k*rows + y, dstStep, x)] = res[k]; + } +} +__constant float c_border[BORDER_SIZE + 1] = { 0.14f, 0.14f, 0.4472f, 0.4472f, 0.4472f, 1.f }; + +__kernel void updateMatrices(__global const float * flowx, int xStep, + __global const float * flowy, int yStep, + const int rows, const int cols, + __global const float * R0, int R0Step, + __global const float * R1, int R1Step, + __global float * M, int mStep) +{ + const int y = get_global_id(1); + const int x = get_global_id(0); + + if (y < rows && y >= 0 && x < cols && x >= 0) + { + float dx = flowx[mad24(y, xStep, x)]; + float dy = flowy[mad24(y, yStep, x)]; + float fx = x + dx; + float fy = y + dy; + + int x1 = convert_int(floor(fx)); + int y1 = convert_int(floor(fy)); + fx -= x1; + fy -= y1; + + float r2, r3, r4, r5, r6; + + if (x1 >= 0 && y1 >= 0 && x1 < cols - 1 && y1 < rows - 1) + { + float a00 = (1.f - fx) * (1.f - fy); + float a01 = fx * (1.f - fy); + float a10 = (1.f - fx) * fy; + float a11 = fx * fy; + + r2 = a00 * R1[mad24(y1, R1Step, x1)] + + a01 * R1[mad24(y1, R1Step, x1 + 1)] + + a10 * R1[mad24(y1 + 1, R1Step, x1)] + + a11 * R1[mad24(y1 + 1, R1Step, x1 + 1)]; + + r3 = a00 * R1[mad24(rows + y1, R1Step, x1)] + + a01 * R1[mad24(rows + y1, R1Step, x1 + 1)] + + a10 * R1[mad24(rows + y1 + 1, R1Step, x1)] + + a11 * R1[mad24(rows + y1 + 1, R1Step, x1 + 1)]; + + r4 = a00 * R1[mad24(2*rows + y1, R1Step, x1)] + + a01 * R1[mad24(2*rows + y1, R1Step, x1 + 1)] + + a10 * R1[mad24(2*rows + y1 + 1, R1Step, x1)] + + a11 * R1[mad24(2*rows + y1 + 1, R1Step, x1 + 1)]; + + r5 = a00 * R1[mad24(3*rows + y1, R1Step, x1)] + + a01 * R1[mad24(3*rows + y1, R1Step, x1 + 1)] + + a10 * R1[mad24(3*rows + y1 + 1, R1Step, x1)] + + a11 * R1[mad24(3*rows + y1 + 1, R1Step, x1 + 1)]; + + r6 = a00 * R1[mad24(4*rows + y1, R1Step, x1)] + + a01 * R1[mad24(4*rows + y1, R1Step, x1 + 1)] + + a10 * R1[mad24(4*rows + y1 + 1, R1Step, x1)] + + a11 * R1[mad24(4*rows + y1 + 1, R1Step, x1 + 1)]; + + r4 = (R0[mad24(2*rows + y, R0Step, x)] + r4) * 0.5f; + r5 = (R0[mad24(3*rows + y, R0Step, x)] + r5) * 0.5f; + r6 = (R0[mad24(4*rows + y, R0Step, x)] + r6) * 0.25f; + } + else + { + r2 = r3 = 0.f; + r4 = R0[mad24(2*rows + y, R0Step, x)]; + r5 = R0[mad24(3*rows + y, R0Step, x)]; + r6 = R0[mad24(4*rows + y, R0Step, x)] * 0.5f; + } + + r2 = (R0[mad24(y, R0Step, x)] - r2) * 0.5f; + r3 = (R0[mad24(rows + y, R0Step, x)] - r3) * 0.5f; + + r2 += r4*dy + r6*dx; + r3 += r6*dy + r5*dx; + + float scale = + c_border[min(x, BORDER_SIZE)] * + c_border[min(y, BORDER_SIZE)] * + c_border[min(cols - x - 1, BORDER_SIZE)] * + c_border[min(rows - y - 1, BORDER_SIZE)]; + + r2 *= scale; + r3 *= scale; + r4 *= scale; + r5 *= scale; + r6 *= scale; + + M[mad24(y, mStep, x)] = r4*r4 + r6*r6; + M[mad24(rows + y, mStep, x)] = (r4 + r5)*r6; + M[mad24(2*rows + y, mStep, x)] = r5*r5 + r6*r6; + M[mad24(3*rows + y, mStep, x)] = r4*r2 + r6*r3; + M[mad24(4*rows + y, mStep, x)] = r6*r2 + r5*r3; + } +} + +__kernel void boxFilter5(__global const float * src, int srcStep, + __global float * dst, int dstStep, + const int rows, const int cols, + const int ksizeHalf, + __local float * smem) +{ + const int y = get_global_id(1); + const int x = get_global_id(0); + + const float boxAreaInv = 1.f / ((1 + 2*ksizeHalf) * (1 + 2*ksizeHalf)); + const int smw = bdx + 2*ksizeHalf; // shared memory "width" + __local float *row = smem + 5 * ty * smw; + + if (y < rows) + { + // Vertical pass + for (int i = tx; i < bdx + 2*ksizeHalf; i += bdx) + { + int xExt = (int)(bx * bdx) + i - ksizeHalf; + xExt = min(max(xExt, 0), cols - 1); + +#pragma unroll + for (int k = 0; k < 5; ++k) + row[k*smw + i] = src[mad24(k*rows + y, srcStep, xExt)]; + + for (int j = 1; j <= ksizeHalf; ++j) +#pragma unroll + for (int k = 0; k < 5; ++k) + row[k*smw + i] += + src[mad24(k*rows + max(y - j, 0), srcStep, xExt)] + + src[mad24(k*rows + min(y + j, rows - 1), srcStep, xExt)]; + } + } + + barrier(CLK_LOCAL_MEM_FENCE); + + if (y < rows && y >= 0 && x < cols && x >= 0) + { + // Horizontal pass + + row += tx + ksizeHalf; + float res[5]; + +#pragma unroll + for (int k = 0; k < 5; ++k) + res[k] = row[k*smw]; + + for (int i = 1; i <= ksizeHalf; ++i) +#pragma unroll + for (int k = 0; k < 5; ++k) + res[k] += row[k*smw - i] + row[k*smw + i]; + +#pragma unroll + for (int k = 0; k < 5; ++k) + dst[mad24(k*rows + y, dstStep, x)] = res[k] * boxAreaInv; + } +} + +__kernel void updateFlow(__global const float * M, int mStep, + __global float * flowx, int xStep, + __global float * flowy, int yStep, + const int rows, const int cols) +{ + const int y = get_global_id(1); + const int x = get_global_id(0); + + if (y < rows && y >= 0 && x < cols && x >= 0) + { + float g11 = M[mad24(y, mStep, x)]; + float g12 = M[mad24(rows + y, mStep, x)]; + float g22 = M[mad24(2*rows + y, mStep, x)]; + float h1 = M[mad24(3*rows + y, mStep, x)]; + float h2 = M[mad24(4*rows + y, mStep, x)]; + + float detInv = 1.f / (g11*g22 - g12*g12 + 1e-3f); + + flowx[mad24(y, xStep, x)] = (g11*h2 - g12*h1) * detInv; + flowy[mad24(y, yStep, x)] = (g22*h1 - g12*h2) * detInv; + } +} \ No newline at end of file diff --git a/modules/video/src/optflowgf.cpp b/modules/video/src/optflowgf.cpp index 19e9688..f4664ae 100644 --- a/modules/video/src/optflowgf.cpp +++ b/modules/video/src/optflowgf.cpp @@ -41,6 +41,7 @@ //M*/ #include "precomp.hpp" +#include "opencl_kernels.hpp" // // 2D dense optical flow algorithm from the following paper: @@ -52,47 +53,40 @@ namespace cv { static void -FarnebackPolyExp( const Mat& src, Mat& dst, int n, double sigma ) +FarnebackPrepareGaussian(int n, double sigma, float *g, float *xg, float *xxg, + double &ig11, double &ig03, double &ig33, double &ig55) { - int k, x, y; - - CV_Assert( src.type() == CV_32FC1 ); - int width = src.cols; - int height = src.rows; - AutoBuffer kbuf(n*6 + 3), _row((width + n*2)*3); - float* g = kbuf + n; - float* xg = g + n*2 + 1; - float* xxg = xg + n*2 + 1; - float *row = (float*)_row + n*3; - if( sigma < FLT_EPSILON ) sigma = n*0.3; double s = 0.; - for( x = -n; x <= n; x++ ) + for (int x = -n; x <= n; x++) { g[x] = (float)std::exp(-x*x/(2*sigma*sigma)); s += g[x]; } s = 1./s; - for( x = -n; x <= n; x++ ) + for (int x = -n; x <= n; x++) { g[x] = (float)(g[x]*s); xg[x] = (float)(x*g[x]); xxg[x] = (float)(x*x*g[x]); } - Mat_ G = Mat_::zeros(6, 6); + Mat_ G(6, 6); + G.setTo(0); - for( y = -n; y <= n; y++ ) - for( x = -n; x <= n; x++ ) + for (int y = -n; y <= n; y++) + { + for (int x = -n; x <= n; x++) { G(0,0) += g[y]*g[x]; G(1,1) += g[y]*g[x]*x*x; G(3,3) += g[y]*g[x]*x*x*x*x; G(5,5) += g[y]*g[x]*x*x*y*y; } + } //G[0][0] = 1.; G(2,2) = G(0,3) = G(0,4) = G(3,0) = G(4,0) = G(1,1); @@ -107,7 +101,29 @@ FarnebackPolyExp( const Mat& src, Mat& dst, int n, double sigma ) // [ e z ] // [ u ] Mat_ invG = G.inv(DECOMP_CHOLESKY); - double ig11 = invG(1,1), ig03 = invG(0,3), ig33 = invG(3,3), ig55 = invG(5,5); + + ig11 = invG(1,1); + ig03 = invG(0,3); + ig33 = invG(3,3); + ig55 = invG(5,5); +} + +static void +FarnebackPolyExp( const Mat& src, Mat& dst, int n, double sigma ) +{ + int k, x, y; + + CV_Assert( src.type() == CV_32FC1 ); + int width = src.cols; + int height = src.rows; + AutoBuffer kbuf(n*6 + 3), _row((width + n*2)*3); + float* g = kbuf + n; + float* xg = g + n*2 + 1; + float* xxg = xg + n*2 + 1; + float *row = (float*)_row + n*3; + double ig11, ig03, ig33, ig55; + + FarnebackPrepareGaussian(n, sigma, g, xg, xxg, ig11, ig03, ig33, ig55); dst.create( height, width, CV_32FC(5)); @@ -563,10 +579,484 @@ FarnebackUpdateFlow_GaussianBlur( const Mat& _R0, const Mat& _R1, } +namespace cv +{ +class FarnebackOpticalFlow +{ +public: + FarnebackOpticalFlow() + { + numLevels = 5; + pyrScale = 0.5; + fastPyramids = false; + winSize = 13; + numIters = 10; + polyN = 5; + polySigma = 1.1; + flags = 0; + } + + int numLevels; + double pyrScale; + bool fastPyramids; + int winSize; + int numIters; + int polyN; + double polySigma; + int flags; + + void operator ()(const UMat &frame0, const UMat &frame1, UMat &flowx, UMat &flowy) + { + CV_Assert(frame0.channels() == 1 && frame1.channels() == 1); + CV_Assert(frame0.size() == frame1.size()); + CV_Assert(polyN == 5 || polyN == 7); + CV_Assert(!fastPyramids || std::abs(pyrScale - 0.5) < 1e-6); + + const int min_size = 32; + + Size size = frame0.size(); + UMat prevFlowX, prevFlowY, curFlowX, curFlowY; + + flowx.create(size, CV_32F); + flowy.create(size, CV_32F); + UMat flowx0 = flowx; + UMat flowy0 = flowy; + + // Crop unnecessary levels + double scale = 1; + int numLevelsCropped = 0; + for (; numLevelsCropped < numLevels; numLevelsCropped++) + { + scale *= pyrScale; + if (size.width*scale < min_size || size.height*scale < min_size) + break; + } + + frame0.convertTo(frames_[0], CV_32F); + frame1.convertTo(frames_[1], CV_32F); + + if (fastPyramids) + { + // Build Gaussian pyramids using pyrDown() + pyramid0_.resize(numLevelsCropped + 1); + pyramid1_.resize(numLevelsCropped + 1); + pyramid0_[0] = frames_[0]; + pyramid1_[0] = frames_[1]; + for (int i = 1; i <= numLevelsCropped; ++i) + { + pyrDown(pyramid0_[i - 1], pyramid0_[i]); + pyrDown(pyramid1_[i - 1], pyramid1_[i]); + } + } + + setPolynomialExpansionConsts(polyN, polySigma); + + for (int k = numLevelsCropped; k >= 0; k--) + { + scale = 1; + for (int i = 0; i < k; i++) + scale *= pyrScale; + + double sigma = (1./scale - 1) * 0.5; + int smoothSize = cvRound(sigma*5) | 1; + smoothSize = std::max(smoothSize, 3); + + int width = cvRound(size.width*scale); + int height = cvRound(size.height*scale); + + if (fastPyramids) + { + width = pyramid0_[k].cols; + height = pyramid0_[k].rows; + } + + if (k > 0) + { + curFlowX.create(height, width, CV_32F); + curFlowY.create(height, width, CV_32F); + } + else + { + curFlowX = flowx0; + curFlowY = flowy0; + } + + if (prevFlowX.empty()) + { + if (flags & cv::OPTFLOW_USE_INITIAL_FLOW) + { + resize(flowx0, curFlowX, Size(width, height), 0, 0, INTER_LINEAR); + resize(flowy0, curFlowY, Size(width, height), 0, 0, INTER_LINEAR); + multiply(scale, curFlowX, curFlowX); + multiply(scale, curFlowY, curFlowY); + } + else + { + curFlowX.setTo(0); + curFlowY.setTo(0); + } + } + else + { + resize(prevFlowX, curFlowX, Size(width, height), 0, 0, INTER_LINEAR); + resize(prevFlowY, curFlowY, Size(width, height), 0, 0, INTER_LINEAR); + multiply(1./pyrScale, curFlowX, curFlowX); + multiply(1./pyrScale, curFlowY, curFlowY); + } + + UMat M = allocMatFromBuf(5*height, width, CV_32F, M_); + UMat bufM = allocMatFromBuf(5*height, width, CV_32F, bufM_); + UMat R[2] = + { + allocMatFromBuf(5*height, width, CV_32F, R_[0]), + allocMatFromBuf(5*height, width, CV_32F, R_[1]) + }; + + if (fastPyramids) + { + polynomialExpansionOcl(pyramid0_[k], polyN, R[0]); + polynomialExpansionOcl(pyramid1_[k], polyN, R[1]); + } + else + { + UMat blurredFrame[2] = + { + allocMatFromBuf(size.height, size.width, CV_32F, blurredFrame_[0]), + allocMatFromBuf(size.height, size.width, CV_32F, blurredFrame_[1]) + }; + UMat pyrLevel[2] = + { + allocMatFromBuf(height, width, CV_32F, pyrLevel_[0]), + allocMatFromBuf(height, width, CV_32F, pyrLevel_[1]) + }; + + setGaussianBlurKernel(smoothSize, sigma); + + for (int i = 0; i < 2; i++) + { + gaussianBlurOcl(frames_[i], smoothSize/2, blurredFrame[i]); + resize(blurredFrame[i], pyrLevel[i], Size(width, height), INTER_LINEAR); + polynomialExpansionOcl(pyrLevel[i], polyN, R[i]); + } + } + + updateMatricesOcl(curFlowX, curFlowY, R[0], R[1], M); + + if (flags & OPTFLOW_FARNEBACK_GAUSSIAN) + setGaussianBlurKernel(winSize, winSize/2*0.3f); + for (int i = 0; i < numIters; i++) + { + if (flags & OPTFLOW_FARNEBACK_GAUSSIAN) + updateFlow_gaussianBlur(R[0], R[1], curFlowX, curFlowY, M, bufM, winSize, i < numIters-1); + else + updateFlow_boxFilter(R[0], R[1], curFlowX, curFlowY, M, bufM, winSize, i < numIters-1); + } + + prevFlowX = curFlowX; + prevFlowY = curFlowY; + } + + flowx = curFlowX; + flowy = curFlowY; + } + + void releaseMemory() + { + frames_[0].release(); + frames_[1].release(); + pyrLevel_[0].release(); + pyrLevel_[1].release(); + M_.release(); + bufM_.release(); + R_[0].release(); + R_[1].release(); + blurredFrame_[0].release(); + blurredFrame_[1].release(); + pyramid0_.clear(); + pyramid1_.clear(); + } +private: + UMat m_g; + UMat m_xg; + UMat m_xxg; + + double m_igd[4]; + float m_ig[4]; + void setPolynomialExpansionConsts(int n, double sigma) + { + std::vector buf(n*6 + 3); + float* g = &buf[0] + n; + float* xg = g + n*2 + 1; + float* xxg = xg + n*2 + 1; + + FarnebackPrepareGaussian(n, sigma, g, xg, xxg, m_igd[0], m_igd[1], m_igd[2], m_igd[3]); + + cv::Mat t_g(1, n + 1, CV_32FC1, g); t_g.copyTo(m_g); + cv::Mat t_xg(1, n + 1, CV_32FC1, xg); t_xg.copyTo(m_xg); + cv::Mat t_xxg(1, n + 1, CV_32FC1, xxg); t_xxg.copyTo(m_xxg); + + m_ig[0] = static_cast(m_igd[0]); + m_ig[1] = static_cast(m_igd[1]); + m_ig[2] = static_cast(m_igd[2]); + m_ig[3] = static_cast(m_igd[3]); + } +private: + UMat m_gKer; + inline void setGaussianBlurKernel(int smoothSize, double sigma) + { + Mat g = getGaussianKernel(smoothSize, sigma, CV_32F); + Mat gKer(1, smoothSize/2 + 1, CV_32FC1, g.ptr(smoothSize/2)); + gKer.copyTo(m_gKer); + } +private: + UMat frames_[2]; + UMat pyrLevel_[2], M_, bufM_, R_[2], blurredFrame_[2]; + std::vector pyramid0_, pyramid1_; + + static UMat allocMatFromBuf(int rows, int cols, int type, UMat &mat) + { + if (!mat.empty() && mat.type() == type && mat.rows >= rows && mat.cols >= cols) + return mat(Rect(0, 0, cols, rows)); + return mat = UMat(rows, cols, type); + } +private: +#define DIVUP(total, grain) (((total) + (grain) - 1) / (grain)) + + bool gaussianBlurOcl(const UMat &src, int ksizeHalf, UMat &dst) + { +#ifdef ANDROID + size_t localsize[2] = { 128, 1}; +#else + size_t localsize[2] = { 256, 1}; +#endif + size_t globalsize[2] = { src.cols, src.rows}; + int smem_size = (int)((localsize[0] + 2*ksizeHalf) * sizeof(float)); + ocl::Kernel kernel; + if (!kernel.create("gaussianBlur", cv::ocl::video::optical_flow_farneback_oclsrc, "")) + return false; + + CV_Assert(dst.size() == src.size()); + int idxArg = 0; + idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadOnly(src)); + idxArg = kernel.set(idxArg, (int)(src.step / src.elemSize())); + idxArg = kernel.set(idxArg, ocl::KernelArg::PtrWriteOnly(dst)); + idxArg = kernel.set(idxArg, (int)(dst.step / dst.elemSize())); + idxArg = kernel.set(idxArg, dst.rows); + idxArg = kernel.set(idxArg, dst.cols); + idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadOnly(m_gKer)); + idxArg = kernel.set(idxArg, (int)ksizeHalf); + idxArg = kernel.set(idxArg, (void *)NULL, smem_size); + return kernel.run(2, globalsize, localsize, false); + } + bool gaussianBlur5Ocl(const UMat &src, int ksizeHalf, UMat &dst) + { + int height = src.rows / 5; +#ifdef ANDROID + size_t localsize[2] = { 128, 1}; +#else + size_t localsize[2] = { 256, 1}; +#endif + size_t globalsize[2] = { src.cols, height}; + int smem_size = (int)((localsize[0] + 2*ksizeHalf) * 5 * sizeof(float)); + ocl::Kernel kernel; + if (!kernel.create("gaussianBlur5", cv::ocl::video::optical_flow_farneback_oclsrc, "")) + return false; + + int idxArg = 0; + idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadOnly(src)); + idxArg = kernel.set(idxArg, (int)(src.step / src.elemSize())); + idxArg = kernel.set(idxArg, ocl::KernelArg::PtrWriteOnly(dst)); + idxArg = kernel.set(idxArg, (int)(dst.step / dst.elemSize())); + idxArg = kernel.set(idxArg, height); + idxArg = kernel.set(idxArg, src.cols); + idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadOnly(m_gKer)); + idxArg = kernel.set(idxArg, (int)ksizeHalf); + idxArg = kernel.set(idxArg, (void *)NULL, smem_size); + return kernel.run(2, globalsize, localsize, false); + } + bool polynomialExpansionOcl(const UMat &src, int polyN, UMat &dst) + { +#ifdef ANDROID + size_t localsize[2] = { 128, 1}; +#else + size_t localsize[2] = { 256, 1}; +#endif + size_t globalsize[2] = { DIVUP(src.cols, localsize[0] - 2*polyN) * localsize[0], src.rows}; + + const cv::ocl::Device &device = cv::ocl::Device::getDefault(); + int useDouble = (0 != device.doubleFPConfig()); + + cv::String build_options = cv::format("-D polyN=%d -D USE_DOUBLE=%d", polyN, useDouble ? 1 : 0); + ocl::Kernel kernel; + if (!kernel.create("polynomialExpansion", cv::ocl::video::optical_flow_farneback_oclsrc, build_options)) + return false; + + int smem_size = (int)(3 * localsize[0] * sizeof(float)); + int idxArg = 0; + idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadOnly(src)); + idxArg = kernel.set(idxArg, (int)(src.step / src.elemSize())); + idxArg = kernel.set(idxArg, ocl::KernelArg::PtrWriteOnly(dst)); + idxArg = kernel.set(idxArg, (int)(dst.step / dst.elemSize())); + idxArg = kernel.set(idxArg, src.rows); + idxArg = kernel.set(idxArg, src.cols); + idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadOnly(m_g)); + idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadOnly(m_xg)); + idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadOnly(m_xxg)); + idxArg = kernel.set(idxArg, (void *)NULL, smem_size); + if (useDouble) + idxArg = kernel.set(idxArg, (void *)m_igd, 4 * sizeof(double)); + else + idxArg = kernel.set(idxArg, (void *)m_ig, 4 * sizeof(float)); + return kernel.run(2, globalsize, localsize, false); + } + bool boxFilter5Ocl(const UMat &src, int ksizeHalf, UMat &dst) + { + int height = src.rows / 5; +#ifdef ANDROID + size_t localsize[2] = { 128, 1}; +#else + size_t localsize[2] = { 256, 1}; +#endif + size_t globalsize[2] = { src.cols, height}; + + ocl::Kernel kernel; + if (!kernel.create("boxFilter5", cv::ocl::video::optical_flow_farneback_oclsrc, "")) + return false; + + int smem_size = (int)((localsize[0] + 2*ksizeHalf) * 5 * sizeof(float)); + + int idxArg = 0; + idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadOnly(src)); + idxArg = kernel.set(idxArg, (int)(src.step / src.elemSize())); + idxArg = kernel.set(idxArg, ocl::KernelArg::PtrWriteOnly(dst)); + idxArg = kernel.set(idxArg, (int)(dst.step / dst.elemSize())); + idxArg = kernel.set(idxArg, height); + idxArg = kernel.set(idxArg, src.cols); + idxArg = kernel.set(idxArg, (int)ksizeHalf); + idxArg = kernel.set(idxArg, (void *)NULL, smem_size); + return kernel.run(2, globalsize, localsize, false); + } + + bool updateFlowOcl(const UMat &M, UMat &flowx, UMat &flowy) + { +#ifdef ANDROID + size_t localsize[2] = { 32, 4}; +#else + size_t localsize[2] = { 32, 8}; +#endif + size_t globalsize[2] = { flowx.cols, flowx.rows}; + + ocl::Kernel kernel; + if (!kernel.create("updateFlow", cv::ocl::video::optical_flow_farneback_oclsrc, "")) + return false; + + int idxArg = 0; + idxArg = kernel.set(idxArg, ocl::KernelArg::PtrWriteOnly(M)); + idxArg = kernel.set(idxArg, (int)(M.step / M.elemSize())); + idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadOnly(flowx)); + idxArg = kernel.set(idxArg, (int)(flowx.step / flowx.elemSize())); + idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadOnly(flowy)); + idxArg = kernel.set(idxArg, (int)(flowy.step / flowy.elemSize())); + idxArg = kernel.set(idxArg, (int)flowy.rows); + idxArg = kernel.set(idxArg, (int)flowy.cols); + return kernel.run(2, globalsize, localsize, false); + } + bool updateMatricesOcl(const UMat &flowx, const UMat &flowy, const UMat &R0, const UMat &R1, UMat &M) + { +#ifdef ANDROID + size_t localsize[2] = { 32, 4}; +#else + size_t localsize[2] = { 32, 8}; +#endif + size_t globalsize[2] = { flowx.cols, flowx.rows}; + + ocl::Kernel kernel; + if (!kernel.create("updateMatrices", cv::ocl::video::optical_flow_farneback_oclsrc, "")) + return false; + + int idxArg = 0; + idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadOnly(flowx)); + idxArg = kernel.set(idxArg, (int)(flowx.step / flowx.elemSize())); + idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadOnly(flowy)); + idxArg = kernel.set(idxArg, (int)(flowy.step / flowy.elemSize())); + idxArg = kernel.set(idxArg, (int)flowx.rows); + idxArg = kernel.set(idxArg, (int)flowx.cols); + idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadOnly(R0)); + idxArg = kernel.set(idxArg, (int)(R0.step / R0.elemSize())); + idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadOnly(R1)); + idxArg = kernel.set(idxArg, (int)(R1.step / R1.elemSize())); + idxArg = kernel.set(idxArg, ocl::KernelArg::PtrWriteOnly(M)); + idxArg = kernel.set(idxArg, (int)(M.step / M.elemSize())); + return kernel.run(2, globalsize, localsize, false); + } + + void updateFlow_boxFilter( + const UMat& R0, const UMat& R1, UMat& flowx, UMat &flowy, + UMat& M, UMat &bufM, int blockSize, bool updateMatrices) + { + boxFilter5Ocl(M, blockSize/2, bufM); + swap(M, bufM); + updateFlowOcl(M, flowx, flowy); + if (updateMatrices) + updateMatricesOcl(flowx, flowy, R0, R1, M); + } + void updateFlow_gaussianBlur( + const UMat& R0, const UMat& R1, UMat& flowx, UMat& flowy, + UMat& M, UMat &bufM, int blockSize, bool updateMatrices) + { + gaussianBlur5Ocl(M, blockSize/2, bufM); + swap(M, bufM); + updateFlowOcl(M, flowx, flowy); + if (updateMatrices) + updateMatricesOcl(flowx, flowy, R0, R1, M); + } +}; + +static bool ocl_calcOpticalFlowFarneback( InputArray _prev0, InputArray _next0, + InputOutputArray _flow0, double pyr_scale, int levels, int winsize, + int iterations, int poly_n, double poly_sigma, int flags ) +{ + if ((5 != poly_n) && (7 != poly_n)) + return false; + if (_next0.size() != _prev0.size()) + return false; + int typePrev = _prev0.type(); + int typeNext = _next0.type(); + if ((1 != CV_MAT_CN(typePrev)) || (1 != CV_MAT_CN(typeNext))) + return false; + + FarnebackOpticalFlow opticalFlow; + opticalFlow.numLevels = levels; + opticalFlow.pyrScale = pyr_scale; + opticalFlow.fastPyramids= false; + opticalFlow.winSize = winsize; + opticalFlow.numIters = iterations; + opticalFlow.polyN = poly_n; + opticalFlow.polySigma = poly_sigma; + opticalFlow.flags = flags; + + std::vector flowar; + if (!_flow0.empty()) + split(_flow0, flowar); + else + { + flowar.push_back(UMat()); + flowar.push_back(UMat()); + } + opticalFlow(_prev0.getUMat(), _next0.getUMat(), flowar[0], flowar[1]); + merge(flowar, _flow0); + return true; +} +} + void cv::calcOpticalFlowFarneback( InputArray _prev0, InputArray _next0, InputOutputArray _flow0, double pyr_scale, int levels, int winsize, int iterations, int poly_n, double poly_sigma, int flags ) { + bool use_opencl = ocl::useOpenCL() && _flow0.isUMat(); + if( use_opencl && ocl_calcOpticalFlowFarneback(_prev0, _next0, _flow0, pyr_scale, levels, winsize, iterations, poly_n, poly_sigma, flags)) + return; + Mat prev0 = _prev0.getMat(), next0 = _next0.getMat(); const int min_size = 32; const Mat* img[2] = { &prev0, &next0 }; diff --git a/modules/video/src/precomp.hpp b/modules/video/src/precomp.hpp index 43ff772..ba0c931 100644 --- a/modules/video/src/precomp.hpp +++ b/modules/video/src/precomp.hpp @@ -46,6 +46,7 @@ #include "opencv2/video.hpp" #include "opencv2/core/utility.hpp" #include "opencv2/core/private.hpp" +#include "opencv2/core/ocl.hpp" #ifdef HAVE_TEGRA_OPTIMIZATION #include "opencv2/video/video_tegra.hpp" diff --git a/modules/video/test/ocl/test_optflow_farneback.cpp b/modules/video/test/ocl/test_optflow_farneback.cpp new file mode 100644 index 0000000..bb35758 --- /dev/null +++ b/modules/video/test/ocl/test_optflow_farneback.cpp @@ -0,0 +1,110 @@ +/*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) 2010-2012, Institute Of Software Chinese Academy Of Science, all rights reserved. +// Copyright (C) 2010-2012, Advanced Micro Devices, Inc., all rights reserved. +// Copyright (C) 2010-2012, Multicoreware, 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 "test_precomp.hpp" +#include "opencv2/ts/ocl_test.hpp" + +#ifdef HAVE_OPENCL + +namespace cvtest { +namespace ocl { + +///////////////////////////////////////////////////////////////////////////////////////////////// +// FarnebackOpticalFlow +namespace +{ + IMPLEMENT_PARAM_CLASS(PyrScale, double) + IMPLEMENT_PARAM_CLASS(PolyN, int) + CV_FLAGS(FarnebackOptFlowFlags, 0, OPTFLOW_FARNEBACK_GAUSSIAN) +} + +PARAM_TEST_CASE(FarnebackOpticalFlow, PyrScale, PolyN, FarnebackOptFlowFlags) +{ + int numLevels; + int winSize; + int numIters; + double pyrScale; + int polyN; + int flags; + + virtual void SetUp() + { + numLevels = 5; + winSize = 13; + numIters = 10; + pyrScale = GET_PARAM(0); + polyN = GET_PARAM(1); + flags = GET_PARAM(2); + } +}; + +OCL_TEST_P(FarnebackOpticalFlow, Accuracy) +{ + cv::Mat frame0 = readImage("optflow/RubberWhale1.png", cv::IMREAD_GRAYSCALE); + ASSERT_FALSE(frame0.empty()); + + cv::Mat frame1 = readImage("optflow/RubberWhale2.png", cv::IMREAD_GRAYSCALE); + ASSERT_FALSE(frame1.empty()); + + double polySigma = polyN <= 5 ? 1.1 : 1.5; + + cv::Mat flow; cv::UMat uflow; + OCL_OFF(cv::calcOpticalFlowFarneback(frame0, frame1, flow, pyrScale, numLevels, winSize, numIters, polyN, polySigma, flags)); + OCL_ON(cv::calcOpticalFlowFarneback(frame0, frame1, uflow, pyrScale, numLevels, winSize, numIters, polyN, polySigma, flags)); + + EXPECT_MAT_SIMILAR(flow, uflow, 0.1) +} + + +OCL_INSTANTIATE_TEST_CASE_P(Video, FarnebackOpticalFlow, + Combine( + Values(PyrScale(0.3), PyrScale(0.5), PyrScale(0.8)), + Values(PolyN(5), PolyN(7)), + Values(FarnebackOptFlowFlags(0), FarnebackOptFlowFlags(cv::OPTFLOW_FARNEBACK_GAUSSIAN)) + ) + ); + + +} } // namespace cvtest::ocl + +#endif // HAVE_OPENCL \ No newline at end of file -- 2.7.4