什么是从cv垫中删除行或col的最佳方法

Question

假设我有一个mat对象,如下所示:

mat = 
   [75, 97, 66, 95, 15, 22;
    24, 21, 71, 72, 34, 66;
    21, 69, 88, 72, 64, 1;
    26, 47, 26, 40, 95, 24;
    70, 37, 9, 83, 16, 83];

我想从它删除一行说第二行有这样的垫子:

 [75, 97, 66, 95, 15, 22;
 21, 69, 88, 72, 64, 1;
 26, 47, 26, 40, 95, 24;
 70, 37, 9, 83, 16, 83]

或删除col col say col 3:

[75, 97,  95, 15, 22;
 24, 21,  72, 34, 66;
 21, 69,  72, 64, 1;
 26, 47,  40, 95, 24;
 70, 37,  83, 16, 83]

最快的方法是什么？我可以将矩阵分解为ROI,然后将它们相互合并,但有没有更好的方法？

Answer 1

我测试了两种方法:

1. 使用cv::Rect和cv::Mat::copyTo:

   // Removing a row
   cv::Mat matIn;    // Matrix of which a row will be deleted.
   int row;          // Row to delete.
   int col;          // Column to delete.
   cv::Mat matOut;   // Result: matIn less that one row.
   
   if ( row > 0 ) // Copy everything above that one row.
   {
       cv::Rect rect( 0, 0, size.width, row );
       matIn( rect ).copyTo( matOut( rect ) );
   }
   
   if ( row < size.height - 1 ) // Copy everything below that one row.
   {
       cv::Rect rect1( 0, row + 1, size.width, size.height - row - 1 );
       cv::Rect rect2( 0, row, size.width, size.height - row - 1 );
       matIn( rect1 ).copyTo( matOut( rect2 ) );
   }
   
   // Removing a column
   if ( col > 0 ) // Copy everything left of that one column.
   {
       cv::Rect rect( 0, 0, col, size.height );
       matIn( rect ).copyTo( matOut( rect ) );
   }
   
   if ( col < size.width - 1 ) // Copy everything right of that one column.
   {
       cv::Rect rect1( col + 1, 0, size.width - col - 1, size.height );
       cv::Rect rect2( col,     0, size.width - col - 1, size.height );
       matIn( rect1 ).copyTo( matOut( rect2 ) );
   }
   

2. 使用std::memcpy和cv::Mat::data:

   // Removing a row
   int rowSizeInBytes = size.width * sizeof( T );
   
   if ( row > 0 )
   {
       int numRows  = row;
       int numBytes = rowSizeInBytes * numRows;
       std::memcpy( matOut.data, matIn.data, numBytes );
   }
   
   if ( row < size.height - 1 )
   {
       int matOutOffset = rowSizeInBytes * row;
       int matInOffset  = matOutOffset + rowSizeInBytes;
   
       int numRows  = size.height - ( row + 1 );
       int numBytes = rowSizeInBytes * numRows;
       std::memcpy( matOut.data + matOutOffset , matIn.data + matInOffset, numBytes );
   }
   
   // Removing a column
   int rowInInBytes  = size.width * sizeof( T );
   int rowOutInBytes = ( size.width - 1 ) * sizeof( T );
   
   if ( col > 0 )
   {
       int matInOffset = 0;
       int matOutOffset = 0;
       int numCols = col;
       int numBytes = numCols * sizeof( T );
   
       for ( int y = 0; y < size.height; ++y )
       {
           std::memcpy( matOut.data + matOutOffset, matIn.data + matInOffset, numBytes );
   
           matInOffset  += rowInInBytes;
           matOutOffset += rowOutInBytes;
       }
   }
   
   if ( col < size.width - 1 )
   {
       int matInOffset = ( col + 1 ) * sizeof( T );
       int matOutOffset = col * sizeof( T );
       int numCols = size.width - ( col + 1 );
       int numBytes = numCols * sizeof( T );
   
       for ( int y = 0; y < size.height; ++y )
       {
           std::memcpy( matOut.data + matOutOffset, matIn.data + matInOffset, numBytes );
   
           matInOffset  += rowInInBytes;
           matOutOffset += rowOutInBytes;
       }
   }
   

第一种方法的定时测试显示:

Removed:      row
Method:       cv::Rect + cv::Mat::copyTo()
Iterations:   10000
Size:         [500 x 500]
Best time:    67ms
Worst time:   526ms
Average time: 70.9061ms
Median time:  70ms

Removed:      column
Method:       cv::Rect + cv::Mat::copyTo()
Iterations:   10000
Size:         [500 x 500]
Best time:    64ms
Worst time:   284ms
Average time: 80.3893ms
Median time:  79ms

对于第二种方法:

Removed:      row
Method:       std::memcpy and/or for-loop
Iterations:   10000
Size:         [500 x 500]
Best time:    31ms
Worst time:   444ms
Average time: 68.9445ms
Median time:  68ms

Removed:      column
Method:       std::memcpy and/or for-loop
Iterations:   10000
Size:         [500 x 500]
Best time:    49ms
Worst time:   122ms
Average time: 79.3948ms
Median time:  78ms

因此,考虑到接近的时序结果和短的实现,第一种方法似乎更合适.我在github上发布了一个最小的工作示例,以验证此测试的结果.