diff --git a/modules/cudaarithm/src/core.cpp b/modules/cudaarithm/src/core.cpp
index ac01afc7f0..aefcfc5af2 100644
--- a/modules/cudaarithm/src/core.cpp
+++ b/modules/cudaarithm/src/core.cpp
@@ -163,8 +163,10 @@ void cv::cuda::flip(InputArray _src, OutputArray _dst, int flipCode, Stream& str
 
     _dst.create(src.size(), src.type());
     GpuMat dst = getOutputMat(_dst, src.size(), src.type(), stream);
+    if (src.data == dst.data && ((src.cols & 1) == 1 || (src.rows & 1) == 1))
+        CV_Error(Error::BadROISize, "In-place version of flip only accepts even width/height");
 
-    if (src.refcount != dst.refcount)
+    if (src.data != dst.data)
         funcs[src.depth()][src.channels() - 1](src, dst, flipCode, StreamAccessor::getStream(stream));
     else // in-place
         ifuncs[src.depth()][src.channels() - 1](src, flipCode, StreamAccessor::getStream(stream));
diff --git a/modules/cudaarithm/test/test_core.cpp b/modules/cudaarithm/test/test_core.cpp
index bc8f3737e5..04b144ff37 100644
--- a/modules/cudaarithm/test/test_core.cpp
+++ b/modules/cudaarithm/test/test_core.cpp
@@ -281,6 +281,8 @@ CUDA_TEST_P(Flip, Accuracy)
 
 CUDA_TEST_P(Flip, AccuracyInplace)
 {
+    size.width  = (size.width  >> 1) << 1; // in-place version only accepts even number
+    size.height = (size.height >> 1) << 1; // in-place version only accepts even number
     cv::Mat src = randomMat(size, type);
 
     cv::cuda::GpuMat srcDst = loadMat(src, useRoi);