Merge pull request #22798 from lamm45:distransform-large

lamm45 · web-flow · commit ddcbd2cc268f · 2023-06-19T15:11:01.000+03:00
Fix distransform to work with large images #22798 This attempts to fix the following bug which was caused by storing squares of large integers into 32-bit floating point variables: #22732 ### Pull Request Readiness Checklist See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request - [x] I agree to contribute to the project under Apache 2 License. - [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV - [x] The PR is proposed to the proper branch - [x] There is a reference to the original bug report and related work - [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable Patch to opencv_extra has the same branch name. - [ ] The feature is well documented and sample code can be built with the project CMake
diff --git a/modules/imgproc/src/distransform.cpp b/modules/imgproc/src/distransform.cpp
@@ -448,7 +448,7 @@ static void getDistanceTransformMask( int maskType, float *metrics )
 
 struct DTColumnInvoker : ParallelLoopBody
 {
-    DTColumnInvoker( const Mat* _src, Mat* _dst, const int* _sat_tab, const float* _sqr_tab)
+    DTColumnInvoker( const Mat* _src, Mat* _dst, const int* _sat_tab, const int* _sqr_tab)
     {
         src = _src;
         dst = _dst;
@@ -481,20 +481,20 @@ struct DTColumnInvoker : ParallelLoopBody
             {
                 dist = dist + 1 - sat_tab[dist - d[j]];
                 d[j] = dist;
-                dptr[0] = sqr_tab[dist];
+                dptr[0] = (float)sqr_tab[dist];
             }
         }
     }
 
     const Mat* src;
     Mat* dst;
     const int* sat_tab;
-    const float* sqr_tab;
+    const int* sqr_tab;
 };
 
 struct DTRowInvoker : ParallelLoopBody
 {
-    DTRowInvoker( Mat* _dst, const float* _sqr_tab, const float* _inv_tab )
+    DTRowInvoker( Mat* _dst, const int* _sqr_tab, const float* _inv_tab )
     {
         dst = _dst;
         sqr_tab = _sqr_tab;
@@ -529,7 +529,7 @@ struct DTRowInvoker : ParallelLoopBody
                 for(;;k--)
                 {
                     p = v[k];
-                    float s = (fq + sqr_tab[q] - d[p] - sqr_tab[p])*inv_tab[q - p];
+                    float s = (fq - d[p] + (sqr_tab[q]-sqr_tab[p]))*inv_tab[q - p];
                     if( s > z[k] )
                     {
                         k++;
@@ -552,28 +552,28 @@ struct DTRowInvoker : ParallelLoopBody
     }
 
     Mat* dst;
-    const float* sqr_tab;
+    const int* sqr_tab;
     const float* inv_tab;
 };
 
 static void
 trueDistTrans( const Mat& src, Mat& dst )
 {
-    const float inf = 1e15f;
+    const int inf = INT_MAX;
 
     CV_Assert( src.size() == dst.size() );
 
     CV_Assert( src.type() == CV_8UC1 && dst.type() == CV_32FC1 );
     int i, m = src.rows, n = src.cols;
 
-    cv::AutoBuffer<uchar> _buf(std::max(m*2*sizeof(float) + (m*3+1)*sizeof(int), n*2*sizeof(float)));
+    cv::AutoBuffer<uchar> _buf(std::max(m*2*sizeof(int) + (m*3+1)*sizeof(int), n*2*sizeof(float)));
     // stage 1: compute 1d distance transform of each column
-    float* sqr_tab = (float*)_buf.data();
+    int* sqr_tab = (int*)_buf.data();
     int* sat_tab = cv::alignPtr((int*)(sqr_tab + m*2), sizeof(int));
     int shift = m*2;
 
     for( i = 0; i < m; i++ )
-        sqr_tab[i] = (float)(i*i);
+        sqr_tab[i] = i*i;
     for( i = m; i < m*2; i++ )
         sqr_tab[i] = inf;
     for( i = 0; i < shift; i++ )
@@ -584,13 +584,14 @@ trueDistTrans( const Mat& src, Mat& dst )
     cv::parallel_for_(cv::Range(0, n), cv::DTColumnInvoker(&src, &dst, sat_tab, sqr_tab), src.total()/(double)(1<<16));
 
     // stage 2: compute modified distance transform for each row
-    float* inv_tab = sqr_tab + n;
+    float* inv_tab = (float*)sqr_tab + n;
 
-    inv_tab[0] = sqr_tab[0] = 0.f;
+    inv_tab[0] = 0.f;
+    sqr_tab[0] = 0;
     for( i = 1; i < n; i++ )
     {
         inv_tab[i] = (float)(0.5/i);
-        sqr_tab[i] = (float)(i*i);
+        sqr_tab[i] = i*i;
     }
 
     cv::parallel_for_(cv::Range(0, m), cv::DTRowInvoker(&dst, sqr_tab, inv_tab));
@@ -752,7 +753,9 @@ void cv::distanceTransform( InputArray _src, OutputArray _dst, OutputArray _labe
         CV_IPP_CHECK()
         {
 #if IPP_DISABLE_PERF_TRUE_DIST_MT
-            if(cv::getNumThreads()<=1 || (src.total()<(int)(1<<14)))
+            // IPP uses floats, but 4097 cannot be squared into a float
+            if((cv::getNumThreads()<=1 || (src.total()<(int)(1<<14))) &&
+                src.rows < 4097 && src.cols < 4097)
 #endif
             {
                 IppStatus status;
diff --git a/modules/imgproc/test/test_distancetransform.cpp b/modules/imgproc/test/test_distancetransform.cpp
@@ -302,4 +302,46 @@ BIGDATA_TEST(Imgproc_DistanceTransform, large_image_12218)
     EXPECT_EQ(nz, (size.height*size.width / 2));
 }
 
+TEST(Imgproc_DistanceTransform, wide_image_22732)
+{
+    Mat src = Mat::zeros(1, 4099, CV_8U); // 4099 or larger used to be bad
+    Mat dist(src.rows, src.cols, CV_32F);
+    distanceTransform(src, dist, DIST_L2, DIST_MASK_PRECISE, CV_32F);
+    int nz = countNonZero(dist);
+    EXPECT_EQ(nz, 0);
+}
+
+TEST(Imgproc_DistanceTransform, large_square_22732)
+{
+    Mat src = Mat::zeros(8000, 8005, CV_8U), dist;
+    distanceTransform(src, dist, DIST_L2, DIST_MASK_PRECISE, CV_32F);
+    int nz = countNonZero(dist);
+    EXPECT_EQ(dist.size(), src.size());
+    EXPECT_EQ(dist.type(), CV_32F);
+    EXPECT_EQ(nz, 0);
+
+    Point p0(src.cols-1, src.rows-1);
+    src.setTo(1);
+    src.at<uchar>(p0) = 0;
+    distanceTransform(src, dist, DIST_L2, DIST_MASK_PRECISE, CV_32F);
+    EXPECT_EQ(dist.size(), src.size());
+    EXPECT_EQ(dist.type(), CV_32F);
+    bool first = true;
+    int nerrs = 0;
+    for (int y = 0; y < dist.rows; y++)
+        for (int x = 0; x < dist.cols; x++) {
+            float d = dist.at<float>(y, x);
+            double dx = (double)(x - p0.x), dy = (double)(y - p0.y);
+            float d0 = (float)sqrt(dx*dx + dy*dy);
+            if (std::abs(d0 - d) > 1) {
+                if (first) {
+                    printf("y=%d, x=%d. dist_ref=%.2f, dist=%.2f\n", y, x, d0, d);
+                    first = false;
+                }
+                nerrs++;
+            }
+        }
+    EXPECT_EQ(0, nerrs) << "reference distance map is different from computed one at " << nerrs << " pixels\n";
+}
+
 }} // namespace
