Merge remote-tracking branch 'upstream/3.4' into merge-3.4

diff --cc modules/dnn/src/dnn.cpp
index f9cbaf4,f650a71..4e65c38
--- 1/modules/dnn/src/dnn.cpp
--- 2/modules/dnn/src/dnn.cpp
+++ b/modules/dnn/src/dnn.cpp
@@@ -2836,11 -2602,7 +2836,13 @@@ struct Net::Impl : public detail::NetIm
                              OpenCLBackendWrapper::update(ld.outputBlobsWrappers, umats);
                          }
  #endif
-                         Range chrange[] = { Range::all(), Range::all(), Range::all(), Range::all() };
++
 +#ifdef HAVE_CUDA
 +                        if (preferableBackend == DNN_BACKEND_CUDA)
 +                            ld.outputBlobsWrappers[0] = wrap(output);
 +#endif
+                         std::vector<Range> chrange(output.dims, Range::all());
++
                          int ofs = 0;
                          for( i = 0; i < ninputs; i++ )
                          {

diff --cc modules/stitching/src/exposure_compensate.cpp
index c204af2,189b2b4..7213349
--- 1/modules/stitching/src/exposure_compensate.cpp
--- 2/modules/stitching/src/exposure_compensate.cpp
+++ b/modules/stitching/src/exposure_compensate.cpp
@@@ -396,30 -254,33 +396,34 @@@ void BlocksCompensator::feed(const std:
          }
      }
  
 -    GainCompensator compensator;
 -    compensator.feed(block_corners, block_images, block_masks);
 -    std::vector<double> gains = compensator.gains();
 -    gain_maps_.resize(num_images);
 +    if (getUpdateGain() || int(gain_maps_.size()) != num_images)
 +    {
 +        Compensator compensator;
 +        compensator.setNrFeeds(getNrFeeds());
 +        compensator.feed(block_corners, block_images, block_masks);
  
 -    Mat_<float> ker(1, 3);
 -    ker(0,0) = 0.25; ker(0,1) = 0.5; ker(0,2) = 0.25;
 +        gain_maps_.clear();
 +        gain_maps_.resize(num_images);
  
 -    int bl_idx = 0;
 -    for (int img_idx = 0; img_idx < num_images; ++img_idx)
 -    {
 -        Size bl_per_img = bl_per_imgs[img_idx];
 -        gain_maps_[img_idx].create(bl_per_img, CV_32F);
 +        Mat_<float> ker(1, 3);
 +        ker(0, 0) = 0.25; ker(0, 1) = 0.5; ker(0, 2) = 0.25;
  
 +        int bl_idx = 0;
 +        for (int img_idx = 0; img_idx < num_images; ++img_idx)
          {
 -            Mat_<float> gain_map = gain_maps_[img_idx].getMat(ACCESS_WRITE);
 -            for (int by = 0; by < bl_per_img.height; ++by)
 -                for (int bx = 0; bx < bl_per_img.width; ++bx, ++bl_idx)
 -                    gain_map(by, bx) = static_cast<float>(gains[bl_idx]);
 -        }
 +            Size bl_per_img = bl_per_imgs[img_idx];
 +            UMat gain_map = getGainMap(compensator, bl_idx, bl_per_img);
 +            bl_idx += bl_per_img.width*bl_per_img.height;
 +
 +            for (int i=0; i<nr_gain_filtering_iterations_; ++i)
-                 sepFilter2D(gain_map, gain_map, CV_32F, ker, ker);
++            {
++                UMat tmp;
++                sepFilter2D(gain_map, tmp, CV_32F, ker, ker);
++                swap(gain_map, tmp);
++            }
  
 -        // 2 smooth passes
 -        UMat result;
 -        sepFilter2D(gain_maps_[img_idx], result, CV_32F, ker, ker);
 -        UMat result2;
 -        sepFilter2D(result, result2, CV_32F, ker, ker);
 -        swap(gain_maps_[img_idx], result2);
 +            gain_maps_[img_idx] = gain_map;
 +        }
      }
  }