Merge pull request #4074 from vpisarev:objdetect_fixes

modules/calib3d/test/test_homography.cpp

@@ -662,7 +662,7 @@ TEST(Calib3d_Homography, fromImages)
     std::vector< DMatch > good_matches;
     for( int i = 0; i < descriptors_1.rows; i++ )
     {
-        if( matches[i].distance <= 42 )
+        if( matches[i].distance <= 100 )
             good_matches.push_back( matches[i]);
     }
 
@@ -676,13 +676,32 @@ TEST(Calib3d_Homography, fromImages)
         pointframe2.push_back( keypoints_2[ good_matches[i].trainIdx ].pt );
     }
 
-    Mat inliers;
+    Mat H0, H1, inliers0, inliers1;
-    Mat H = findHomography( pointframe1, pointframe2, RANSAC,3.0,inliers);
+    double min_t0 = DBL_MAX, min_t1 = DBL_MAX;
-    int ninliers = countNonZero(inliers);
+    for( int i = 0; i < 10; i++ )
-    printf("nfeatures1 = %d, nfeatures2=%d, good matches=%d, ninliers=%d\n",
+    {
+        double t = (double)getTickCount();
+        H0 = findHomography( pointframe1, pointframe2, RANSAC, 3.0, inliers0 );
+        t = (double)getTickCount() - t;
+        min_t0 = std::min(min_t0, t);
+    }
+    int ninliers0 = countNonZero(inliers0);
+    for( int i = 0; i < 10; i++ )
+    {
+        double t = (double)getTickCount();
+        H1 = findHomography( pointframe1, pointframe2, RHO, 3.0, inliers1 );
+        t = (double)getTickCount() - t;
+        min_t1 = std::min(min_t1, t);
+    }
+    int ninliers1 = countNonZero(inliers1);
+    double freq = getTickFrequency();
+    printf("nfeatures1 = %d, nfeatures2=%d, matches=%d, ninliers(RANSAC)=%d, "
+           "time(RANSAC)=%.2fmsec, ninliers(RHO)=%d, time(RHO)=%.2fmsec\n",
            (int)keypoints_1.size(), (int)keypoints_2.size(),
-           (int)good_matches.size(), ninliers);
+           (int)good_matches.size(), ninliers0, min_t0*1000./freq, ninliers1, min_t1*1000./freq);
 
-    ASSERT_TRUE(!H.empty());
+    ASSERT_TRUE(!H0.empty());
-    ASSERT_GE(ninliers, 80);
+    ASSERT_GE(ninliers0, 80);
+    ASSERT_TRUE(!H1.empty());
+    ASSERT_GE(ninliers1, 80);
 }

modules/core/test/test_hal_core.cpp

@@ -0,0 +1,195 @@
+/*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) 2013, OpenCV Foundation, 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 OpenCV Foundation 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/hal.hpp"
+
+using namespace cv;
+
+enum
+{
+    HAL_EXP = 0,
+    HAL_LOG = 1,
+    HAL_SQRT = 2
+};
+
+TEST(Core_HAL, mathfuncs)
+{
+    for( int hcase = 0; hcase < 6; hcase++ )
+    {
+        int depth = hcase % 2 == 0 ? CV_32F : CV_64F;
+        double eps = depth == CV_32F ? 1e-5 : 1e-10;
+        int nfunc = hcase / 2;
+        int n = 100;
+
+        Mat src(1, n, depth), dst(1, n, depth), dst0(1, n, depth);
+        randu(src, 1, 10);
+
+        double min_hal_t = DBL_MAX, min_ocv_t = DBL_MAX;
+
+        for( int iter = 0; iter < 10; iter++ )
+        {
+            double t = (double)getTickCount();
+            switch (nfunc)
+            {
+            case HAL_EXP:
+                if( depth == CV_32F )
+                    hal::exp(src.ptr<float>(), dst.ptr<float>(), n);
+                else
+                    hal::exp(src.ptr<double>(), dst.ptr<double>(), n);
+                break;
+            case HAL_LOG:
+                if( depth == CV_32F )
+                    hal::log(src.ptr<float>(), dst.ptr<float>(), n);
+                else
+                    hal::log(src.ptr<double>(), dst.ptr<double>(), n);
+                break;
+            case HAL_SQRT:
+                if( depth == CV_32F )
+                    hal::sqrt(src.ptr<float>(), dst.ptr<float>(), n);
+                else
+                    hal::sqrt(src.ptr<double>(), dst.ptr<double>(), n);
+                break;
+            default:
+                CV_Error(Error::StsBadArg, "unknown function");
+            }
+            t = (double)getTickCount() - t;
+            min_hal_t = std::min(min_hal_t, t);
+
+            t = (double)getTickCount();
+            switch (nfunc)
+            {
+            case HAL_EXP:
+                exp(src, dst0);
+                break;
+            case HAL_LOG:
+                log(src, dst0);
+                break;
+            case HAL_SQRT:
+                pow(src, 0.5, dst0);
+                break;
+            default:
+                CV_Error(Error::StsBadArg, "unknown function");
+            }
+            t = (double)getTickCount() - t;
+            min_ocv_t = std::min(min_ocv_t, t);
+        }
+        EXPECT_LE(norm(dst, dst0, NORM_INF | NORM_RELATIVE), eps);
+
+        double freq = getTickFrequency();
+        printf("%s (N=%d, %s): hal time=%.2fusec, ocv time=%.2fusec\n",
+               (nfunc == HAL_EXP ? "exp" : nfunc == HAL_LOG ? "log" : nfunc == HAL_SQRT ? "sqrt" : "???"),
+               n, (depth == CV_32F ? "f32" : "f64"), min_hal_t*1e6/freq, min_ocv_t*1e6/freq);
+    }
+}
+
+enum
+{
+    HAL_LU = 0,
+    HAL_CHOL = 1
+};
+
+TEST(Core_HAL, mat_decomp)
+{
+    for( int hcase = 0; hcase < 16; hcase++ )
+    {
+        int depth = hcase % 2 == 0 ? CV_32F : CV_64F;
+        int size = (hcase / 2) % 4;
+        size = size == 0 ? 3 : size == 1 ? 4  : size == 2 ? 6 : 15;
+        int nfunc = (hcase / 8);
+        double eps = depth == CV_32F ? 1e-5 : 1e-10;
+
+        if( size == 3 )
+            continue;
+
+        Mat a0(size, size, depth), a(size, size, depth), b(size, 1, depth), x(size, 1, depth), x0(size, 1, depth);
+        randu(a0, -1, 1);
+        a0 = a0*a0.t();
+        randu(b, -1, 1);
+
+        double min_hal_t = DBL_MAX, min_ocv_t = DBL_MAX;
+        size_t asize = size*size*a.elemSize();
+        size_t bsize = size*b.elemSize();
+
+        for( int iter = 0; iter < 10; iter++ )
+        {
+            memcpy(x.ptr(), b.ptr(), bsize);
+            memcpy(a.ptr(), a0.ptr(), asize);
+
+            double t = (double)getTickCount();
+            switch (nfunc)
+            {
+            case HAL_LU:
+                if( depth == CV_32F )
+                    hal::LU(a.ptr<float>(), a.step, size, x.ptr<float>(), x.step, 1);
+                else
+                    hal::LU(a.ptr<double>(), a.step, size, x.ptr<double>(), x.step, 1);
+                break;
+            case HAL_CHOL:
+                if( depth == CV_32F )
+                    hal::Cholesky(a.ptr<float>(), a.step, size, x.ptr<float>(), x.step, 1);
+                else
+                    hal::Cholesky(a.ptr<double>(), a.step, size, x.ptr<double>(), x.step, 1);
+                break;
+            default:
+                CV_Error(Error::StsBadArg, "unknown function");
+            }
+            t = (double)getTickCount() - t;
+            min_hal_t = std::min(min_hal_t, t);
+
+            t = (double)getTickCount();
+            solve(a0, b, x0, (nfunc == HAL_LU ? DECOMP_LU : DECOMP_CHOLESKY));
+            t = (double)getTickCount() - t;
+            min_ocv_t = std::min(min_ocv_t, t);
+        }
+        //std::cout << "x: " << Mat(x.t()) << std::endl;
+        //std::cout << "x0: " << Mat(x0.t()) << std::endl;
+
+        EXPECT_LE(norm(x, x0, NORM_INF | NORM_RELATIVE), eps);
+
+        double freq = getTickFrequency();
+        printf("%s (%d x %d, %s): hal time=%.2fusec, ocv time=%.2fusec\n",
+               (nfunc == HAL_LU ? "LU" : nfunc == HAL_CHOL ? "Cholesky" : "???"),
+               size, size,
+               (depth == CV_32F ? "f32" : "f64"),
+               min_hal_t*1e6/freq, min_ocv_t*1e6/freq);
+    }
+}

modules/imgproc/src/sumpixels.cpp

@@ -318,6 +318,7 @@ static void integral_##suffix( T* src, size_t srcstep, ST* sum, size_t sumstep,
 { integral_(src, srcstep, sum, sumstep, sqsum, sqsumstep, tilted, tiltedstep, size, cn); }
 
 DEF_INTEGRAL_FUNC(8u32s, uchar, int, double)
+DEF_INTEGRAL_FUNC(8u32s32s, uchar, int, int)
 DEF_INTEGRAL_FUNC(8u32f64f, uchar, float, double)
 DEF_INTEGRAL_FUNC(8u64f64f, uchar, double, double)
 DEF_INTEGRAL_FUNC(16u64f64f, ushort, double, double)
@@ -505,6 +506,8 @@ void cv::integral( InputArray _src, OutputArray _sum, OutputArray _sqsum, Output
         func = (IntegralFunc)GET_OPTIMIZED(integral_8u32s);
     else if( depth == CV_8U && sdepth == CV_32S && sqdepth == CV_32F )
         func = (IntegralFunc)integral_8u32s32f;
+    else if( depth == CV_8U && sdepth == CV_32S && sqdepth == CV_32S )
+        func = (IntegralFunc)integral_8u32s32s;
     else if( depth == CV_8U && sdepth == CV_32F && sqdepth == CV_64F )
         func = (IntegralFunc)integral_8u32f64f;
     else if( depth == CV_8U && sdepth == CV_32F && sqdepth == CV_32F )

modules/objdetect/src/cascadedetect.cpp

@@ -627,33 +627,33 @@ void HaarEvaluator::computeChannels(int scaleIdx, InputArray img)
         int sqy = sy + (sqofs / sbufSize.width);
         UMat sum(usbuf, Rect(sx, sy, s.szi.width, s.szi.height));
         UMat sqsum(usbuf, Rect(sx, sqy, s.szi.width, s.szi.height));
-        sqsum.flags = (sqsum.flags & ~UMat::DEPTH_MASK) | CV_32F;
+        sqsum.flags = (sqsum.flags & ~UMat::DEPTH_MASK) | CV_32S;
 
         if (hasTiltedFeatures)
         {
             int sty = sy + (tofs / sbufSize.width);
             UMat tilted(usbuf, Rect(sx, sty, s.szi.width, s.szi.height));
-            integral(img, sum, sqsum, tilted, CV_32S, CV_32F);
+            integral(img, sum, sqsum, tilted, CV_32S, CV_32S);
         }
         else
         {
             UMatData* u = sqsum.u;
-            integral(img, sum, sqsum, noArray(), CV_32S, CV_32F);
+            integral(img, sum, sqsum, noArray(), CV_32S, CV_32S);
-            CV_Assert(sqsum.u == u && sqsum.size() == s.szi && sqsum.type()==CV_32F);
+            CV_Assert(sqsum.u == u && sqsum.size() == s.szi && sqsum.type()==CV_32S);
         }
     }
     else
     {
         Mat sum(s.szi, CV_32S, sbuf.ptr<int>() + s.layer_ofs, sbuf.step);
-        Mat sqsum(s.szi, CV_32F, sum.ptr<int>() + sqofs, sbuf.step);
+        Mat sqsum(s.szi, CV_32S, sum.ptr<int>() + sqofs, sbuf.step);
 
         if (hasTiltedFeatures)
         {
             Mat tilted(s.szi, CV_32S, sum.ptr<int>() + tofs, sbuf.step);
-            integral(img, sum, sqsum, tilted, CV_32S, CV_32F);
+            integral(img, sum, sqsum, tilted, CV_32S, CV_32S);
         }
         else
-            integral(img, sum, sqsum, noArray(), CV_32S, CV_32F);
+            integral(img, sum, sqsum, noArray(), CV_32S, CV_32S);
     }
 }
 
@@ -689,18 +689,23 @@ bool HaarEvaluator::setWindow( Point pt, int scaleIdx )
         return false;
 
     pwin = &sbuf.at<int>(pt) + s.layer_ofs;
-    const float* pq = (const float*)(pwin + sqofs);
+    const int* pq = (const int*)(pwin + sqofs);
     int valsum = CALC_SUM_OFS(nofs, pwin);
-    float valsqsum = CALC_SUM_OFS(nofs, pq);
+    unsigned valsqsum = (unsigned)(CALC_SUM_OFS(nofs, pq));
 
-    double nf = (double)normrect.area() * valsqsum - (double)valsum * valsum;
+    double area = normrect.area();
+    double nf = area * valsqsum - (double)valsum * valsum;
     if( nf > 0. )
+    {
         nf = std::sqrt(nf);
+        varianceNormFactor = (float)(1./nf);
+        return area*varianceNormFactor < 1e-1;
+    }
     else
-        nf = 1.;
+    {
-    varianceNormFactor = (float)(1./nf);
+        varianceNormFactor = 1.f;
-
+        return false;
-    return true;
+    }
 }
 
 
@@ -1402,8 +1407,10 @@ bool CascadeClassifierImpl::Data::read(const FileNode &root)
     else if( featureTypeStr == CC_LBP )
         featureType = FeatureEvaluator::LBP;
     else if( featureTypeStr == CC_HOG )
+    {
         featureType = FeatureEvaluator::HOG;
-
+        CV_Error(Error::StsNotImplemented, "HOG cascade is not supported in 3.0");
+    }
     else
         return false;
 
@@ -1580,6 +1587,43 @@ bool CascadeClassifier::read(const FileNode &root)
     return ok;
 }
 
+void clipObjects(Size sz, std::vector<Rect>& objects,
+                 std::vector<int>* a, std::vector<double>* b)
+{
+    size_t i, j = 0, n = objects.size();
+    Rect win0 = Rect(0, 0, sz.width, sz.height);
+    if(a)
+    {
+        CV_Assert(a->size() == n);
+    }
+    if(b)
+    {
+        CV_Assert(b->size() == n);
+    }
+
+    for( i = 0; i < n; i++ )
+    {
+        Rect r = win0 & objects[i];
+        if( r.area() > 0 )
+        {
+            objects[j] = r;
+            if( i > j )
+            {
+                if(a) a->at(j) = a->at(i);
+                if(b) b->at(j) = b->at(i);
+            }
+            j++;
+        }
+    }
+
+    if( j < n )
+    {
+        objects.resize(j);
+        if(a) a->resize(j);
+        if(b) b->resize(j);
+    }
+}
+
 void CascadeClassifier::detectMultiScale( InputArray image,
                       CV_OUT std::vector<Rect>& objects,
                       double scaleFactor,
@@ -1589,6 +1633,7 @@ void CascadeClassifier::detectMultiScale( InputArray image,
 {
     CV_Assert(!empty());
     cc->detectMultiScale(image, objects, scaleFactor, minNeighbors, flags, minSize, maxSize);
+    clipObjects(image.size(), objects, 0, 0);
 }
 
 void CascadeClassifier::detectMultiScale( InputArray image,
@@ -1601,6 +1646,7 @@ void CascadeClassifier::detectMultiScale( InputArray image,
     CV_Assert(!empty());
     cc->detectMultiScale(image, objects, numDetections,
                          scaleFactor, minNeighbors, flags, minSize, maxSize);
+    clipObjects(image.size(), objects, &numDetections, 0);
 }
 
 void CascadeClassifier::detectMultiScale( InputArray image,
@@ -1616,6 +1662,7 @@ void CascadeClassifier::detectMultiScale( InputArray image,
     cc->detectMultiScale(image, objects, rejectLevels, levelWeights,
                          scaleFactor, minNeighbors, flags,
                          minSize, maxSize, outputRejectLevels);
+    clipObjects(image.size(), objects, &rejectLevels, &levelWeights);
 }
 
 bool CascadeClassifier::isOldFormatCascade() const

modules/objdetect/src/cascadedetect.hpp

@@ -5,6 +5,9 @@
 namespace cv
 {
 
+void clipObjects(Size sz, std::vector<Rect>& objects,
+                 std::vector<int>* a, std::vector<double>* b);
+
 class FeatureEvaluator
 {
 public:

modules/objdetect/src/hog.cpp

@@ -41,6 +41,7 @@
 //M*/
 
 #include "precomp.hpp"
+#include "cascadedetect.hpp"
 #include "opencv2/core/core_c.h"
 #include "opencl_kernels_objdetect.hpp"
 
@@ -1823,7 +1824,9 @@ static bool ocl_detectMultiScale(InputArray _img, std::vector<Rect> &found_locat
             all_candidates.push_back(Rect(Point2d(locations[j]) * scale, scaled_win_size));
     }
     found_locations.assign(all_candidates.begin(), all_candidates.end());
-    cv::groupRectangles(found_locations, (int)group_threshold, 0.2);
+    groupRectangles(found_locations, (int)group_threshold, 0.2);
+    clipObjects(imgSize, found_locations, 0, 0);
+
     return true;
 }
 #endif //HAVE_OPENCL
@@ -1879,6 +1882,7 @@ void HOGDescriptor::detectMultiScale(
         groupRectangles_meanshift(foundLocations, foundWeights, foundScales, finalThreshold, winSize);
     else
         groupRectangles(foundLocations, foundWeights, (int)finalThreshold, 0.2);
+    clipObjects(imgSize, foundLocations, 0, &foundWeights);
 }
 
 void HOGDescriptor::detectMultiScale(InputArray img, std::vector<Rect>& foundLocations,

modules/objdetect/src/opencl/cascadedetect.cl

@@ -160,7 +160,7 @@ void runHaarClassifier(
                 __global const int* psum = psum1;
                 #endif
 
-                __global const float* psqsum = (__global const float*)(psum1 + sqofs);
+                __global const int* psqsum = (__global const int*)(psum1 + sqofs);
                 float sval = (psum[nofs.x] - psum[nofs.y] - psum[nofs.z] + psum[nofs.w])*invarea;
                 float sqval = (psqsum[nofs0.x] - psqsum[nofs0.y] - psqsum[nofs0.z] + psqsum[nofs0.w])*invarea;
                 float nf = (float)normarea * sqrt(max(sqval - sval * sval, 0.f));

modules/objdetect/test/test_cascadeandhog.cpp

@@ -1360,4 +1360,32 @@ TEST(Objdetect_HOGDetector_Strict, accuracy)
         std::vector<float> descriptors;
         reference_hog.compute(image, descriptors);
     }
- }
+}
+
+TEST(Objdetect_CascadeDetector, small_img)
+{
+    String root = cvtest::TS::ptr()->get_data_path() + "cascadeandhog/cascades/";
+    String cascades[] =
+    {
+        root + "haarcascade_frontalface_alt.xml",
+        root + "lbpcascade_frontalface.xml",
+        String()
+    };
+
+    vector<Rect> objects;
+    RNG rng((uint64)-1);
+
+    for( int i = 0; !cascades[i].empty(); i++ )
+    {
+        printf("%d. %s\n", i, cascades[i].c_str());
+        CascadeClassifier cascade(cascades[i]);
+        for( int j = 0; j < 100; j++ )
+        {
+            int width = rng.uniform(1, 100);
+            int height = rng.uniform(1, 100);
+            Mat img(height, width, CV_8U);
+            randu(img, 0, 256);
+            cascade.detectMultiScale(img, objects);
+        }
+    }
+}

samples/python2/facedetect.py

@@ -49,11 +49,12 @@ if __name__ == '__main__':
         rects = detect(gray, cascade)
         vis = img.copy()
         draw_rects(vis, rects, (0, 255, 0))
-        for x1, y1, x2, y2 in rects:
+        if not nested.empty():
-            roi = gray[y1:y2, x1:x2]
+            for x1, y1, x2, y2 in rects:
-            vis_roi = vis[y1:y2, x1:x2]
+                roi = gray[y1:y2, x1:x2]
-            subrects = detect(roi.copy(), nested)
+                vis_roi = vis[y1:y2, x1:x2]
-            draw_rects(vis_roi, subrects, (255, 0, 0))
+                subrects = detect(roi.copy(), nested)
+                draw_rects(vis_roi, subrects, (255, 0, 0))
         dt = clock() - t
 
         draw_str(vis, (20, 20), 'time: %.1f ms' % (dt*1000))