From 603bc58d3bd7bcb7386adc40c094b911a9199d1d Mon Sep 17 00:00:00 2001
From: Samaresh Kumar Singh <ssam3003@gmail.com>
Date: Sat, 8 Nov 2025 13:16:01 -0600
Subject: [PATCH] Fix #27968: Eliminate unnecessary type conversions in GrabCut
 initGMMs

Remove performance bottleneck caused by redundant type conversions in the
initGMMs() function's tight nested loops that iterate over all image pixels.

Changes:
- Changed storage vectors from Vec3f to Vec3b to store pixel data directly
  without intermediate conversion
- Modified kmeans preparation to convert Vec3b data to CV_32FC1 only when
  creating the Mat for clustering (using convertTo instead of per-pixel cast)
- Explicitly convert Vec3b to Vec3d only when calling addSample() method

Performance Impact:
Previously: Vec3b -> Vec3f (in loop) -> Vec3d (implicit in addSample)
Now: Vec3b (in loop) -> Vec3d (explicit, only in addSample call)

This eliminates one unnecessary type conversion per pixel in the tight loop,
reducing computational overhead especially for large images.

Fixes #27968
---
 modules/imgproc/src/grabcut.cpp | 16 +++++++++-------
 1 file changed, 9 insertions(+), 7 deletions(-)

diff --git a/modules/imgproc/src/grabcut.cpp b/modules/imgproc/src/grabcut.cpp
index 358747843e..115f346e99 100644
--- a/modules/imgproc/src/grabcut.cpp
+++ b/modules/imgproc/src/grabcut.cpp
@@ -371,28 +371,30 @@ static void initGMMs( const Mat& img, const Mat& mask, GMM& bgdGMM, GMM& fgdGMM
     const int kMeansType = KMEANS_PP_CENTERS;
 
     Mat bgdLabels, fgdLabels;
-    std::vector<Vec3f> bgdSamples, fgdSamples;
+    std::vector<Vec3b> bgdSamples, fgdSamples;
     Point p;
     for( p.y = 0; p.y < img.rows; p.y++ )
     {
         for( p.x = 0; p.x < img.cols; p.x++ )
         {
             if( mask.at<uchar>(p) == GC_BGD || mask.at<uchar>(p) == GC_PR_BGD )
-                bgdSamples.push_back( (Vec3f)img.at<Vec3b>(p) );
+                bgdSamples.push_back( img.at<Vec3b>(p) );
             else // GC_FGD | GC_PR_FGD
-                fgdSamples.push_back( (Vec3f)img.at<Vec3b>(p) );
+                fgdSamples.push_back( img.at<Vec3b>(p) );
         }
     }
     CV_Assert( !bgdSamples.empty() && !fgdSamples.empty() );
     {
-        Mat _bgdSamples( (int)bgdSamples.size(), 3, CV_32FC1, &bgdSamples[0][0] );
+        Mat _bgdSamples( (int)bgdSamples.size(), 3, CV_8UC1, &bgdSamples[0][0] );
+        _bgdSamples.convertTo(_bgdSamples, CV_32FC1);
         int num_clusters = GMM::componentsCount;
         num_clusters = std::min(num_clusters, (int)bgdSamples.size());
         kmeans( _bgdSamples, num_clusters, bgdLabels,
                 TermCriteria( TermCriteria::MAX_ITER, kMeansItCount, 0.0), 0, kMeansType );
     }
     {
-        Mat _fgdSamples( (int)fgdSamples.size(), 3, CV_32FC1, &fgdSamples[0][0] );
+        Mat _fgdSamples( (int)fgdSamples.size(), 3, CV_8UC1, &fgdSamples[0][0] );
+        _fgdSamples.convertTo(_fgdSamples, CV_32FC1);
         int num_clusters = GMM::componentsCount;
         num_clusters = std::min(num_clusters, (int)fgdSamples.size());
         kmeans( _fgdSamples, num_clusters, fgdLabels,
@@ -401,12 +403,12 @@ static void initGMMs( const Mat& img, const Mat& mask, GMM& bgdGMM, GMM& fgdGMM
 
     bgdGMM.initLearning();
     for( int i = 0; i < (int)bgdSamples.size(); i++ )
-        bgdGMM.addSample( bgdLabels.at<int>(i,0), bgdSamples[i] );
+        bgdGMM.addSample( bgdLabels.at<int>(i,0), Vec3d(bgdSamples[i]) );
     bgdGMM.endLearning();
 
     fgdGMM.initLearning();
     for( int i = 0; i < (int)fgdSamples.size(); i++ )
-        fgdGMM.addSample( fgdLabels.at<int>(i,0), fgdSamples[i] );
+        fgdGMM.addSample( fgdLabels.at<int>(i,0), Vec3d(fgdSamples[i]) );
     fgdGMM.endLearning();
 }