diff --git a/.bazelrc b/.bazelrc
index 5a586f3ca..2834c7012 100644
--- a/.bazelrc
+++ b/.bazelrc
@@ -83,3 +83,13 @@ build:ios_fat --watchos_cpus=armv7k
 build:darwin_x86_64 --apple_platform_type=macos
 build:darwin_x86_64 --macos_minimum_os=10.12
 build:darwin_x86_64 --cpu=darwin_x86_64
+
+# This config refers to building with CUDA available. It does not necessarily
+# mean that we build CUDA op kernels.
+build:using_cuda --define=using_cuda=true
+build:using_cuda --action_env TF_NEED_CUDA=1
+build:using_cuda --crosstool_top=@local_config_cuda//crosstool:toolchain
+
+# This config refers to building CUDA op kernels with nvcc.
+build:cuda --config=using_cuda
+build:cuda --define=using_cuda_nvcc=true
diff --git a/mediapipe/calculators/video/tvl1_optical_flow_calculator.cc b/mediapipe/calculators/video/tvl1_optical_flow_calculator.cc
index c774cfeb1..e6130f8b5 100644
--- a/mediapipe/calculators/video/tvl1_optical_flow_calculator.cc
+++ b/mediapipe/calculators/video/tvl1_optical_flow_calculator.cc
@@ -88,7 +88,7 @@ class Tvl1OpticalFlowCalculator : public CalculatorBase {
   // cv::DenseOpticalFlow is not thread-safe. Invoking multiple
   // DenseOpticalFlow::calc() in parallel may lead to memory corruption or
   // memory leak.
-  std::list<cv::Ptr<cv::DenseOpticalFlow>> tvl1_computers_
+  std::list<cv::Ptr<DenseOpticalFlow>> tvl1_computers_
       ABSL_GUARDED_BY(mutex_);
   absl::Mutex mutex_;
 };
@@ -163,7 +163,7 @@ class Tvl1OpticalFlowCalculator : public CalculatorBase {
 
   // Tries getting an idle DenseOpticalFlow object from the cache. If not,
   // creates a new DenseOpticalFlow.
-  cv::Ptr<cv::DenseOpticalFlow> tvl1_computer;
+  cv::Ptr<DenseOpticalFlow> tvl1_computer;
   {
     absl::MutexLock lock(&mutex_);
     if (!tvl1_computers_.empty()) {
@@ -177,7 +177,18 @@ class Tvl1OpticalFlowCalculator : public CalculatorBase {
 
   flow->Allocate(first.cols, first.rows);
   cv::Mat cv_flow(flow->mutable_flow_data());
+#if defined(CV_CUDA)
+  cv::cuda::GpuMat gpu_first, gpu_second, gpu_flow;
+  gpu_first.upload(first);
+  gpu_second.upload(second);
+  gpu_flow.upload(cv_flow);
+  tvl1_computer->calc(gpu_first, gpu_second, gpu_flow);
+  gpu_first.download(first);
+  gpu_second.download(second);
+  gpu_flow.download(cv_flow);
+#else
   tvl1_computer->calc(first, second, cv_flow);
+#endif
   CHECK_EQ(flow->mutable_flow_data().data, cv_flow.data);
   // Inserts the idle DenseOpticalFlow object back to the cache for reuse.
   {
diff --git a/mediapipe/framework/port/opencv_video_inc.h b/mediapipe/framework/port/opencv_video_inc.h
index dc84bf59b..05fbd2782 100644
--- a/mediapipe/framework/port/opencv_video_inc.h
+++ b/mediapipe/framework/port/opencv_video_inc.h
@@ -84,13 +84,26 @@ inline int fourcc(char c1, char c2, char c3, char c4) {
 #include <opencv2/video.hpp>
 #include <opencv2/videoio.hpp>
 
+//#define CV_CUDA
 #if CV_VERSION_MAJOR == 4 && !defined(MEDIAPIPE_MOBILE)
+#if defined(CV_CUDA)
+#include <opencv2/cudaoptflow.hpp>
+typedef cv::cuda::DenseOpticalFlow DenseOpticalFlow;
+#else
 #include <opencv2/optflow.hpp>
+typedef cv::DenseOpticalFlow DenseOpticalFlow;
+#endif
 
 namespace cv {
+#if defined(CV_CUDA)
+inline Ptr<cuda::DenseOpticalFlow> createOptFlow_DualTVL1() {
+  return cuda::OpticalFlowDual_TVL1::create();
+}
+#else
 inline Ptr<DenseOpticalFlow> createOptFlow_DualTVL1() {
   return optflow::createOptFlow_DualTVL1();
 }
+#endif
 }  // namespace cv
 #endif
 
diff --git a/third_party/opencv_linux.BUILD b/third_party/opencv_linux.BUILD
index 26978d532..026734ab9 100644
--- a/third_party/opencv_linux.BUILD
+++ b/third_party/opencv_linux.BUILD
@@ -23,6 +23,8 @@ cc_library(
             "lib/x86_64-linux-gnu/libopencv_imgproc.so",
             "lib/x86_64-linux-gnu/libopencv_video.so",
             "lib/x86_64-linux-gnu/libopencv_videoio.so",
+            "lib/x86_64-linux-gnu/libopencv_optflow.so",
+            "lib/x86_64-linux-gnu/libopencv_cudaoptflow.so",
         ],
     ),
     hdrs = glob([