diff --git a/apps/annotation/CMakeLists.txt b/apps/annotation/CMakeLists.txt
index 57b133df12..9288e86b42 100644
--- a/apps/annotation/CMakeLists.txt
+++ b/apps/annotation/CMakeLists.txt
@@ -21,7 +21,6 @@ set_target_properties(${the_target} PROPERTIES
                       DEBUG_POSTFIX "${OPENCV_DEBUG_POSTFIX}"
                       ARCHIVE_OUTPUT_DIRECTORY ${LIBRARY_OUTPUT_PATH}
                       RUNTIME_OUTPUT_DIRECTORY ${EXECUTABLE_OUTPUT_PATH}
-                      INSTALL_NAME_DIR lib
                       OUTPUT_NAME "opencv_annotation")
 
 if(ENABLE_SOLUTION_FOLDERS)
diff --git a/apps/createsamples/CMakeLists.txt b/apps/createsamples/CMakeLists.txt
index 24506231e0..a285c69e41 100644
--- a/apps/createsamples/CMakeLists.txt
+++ b/apps/createsamples/CMakeLists.txt
@@ -23,7 +23,6 @@ set_target_properties(${the_target} PROPERTIES
                       DEBUG_POSTFIX "${OPENCV_DEBUG_POSTFIX}"
                       ARCHIVE_OUTPUT_DIRECTORY ${LIBRARY_OUTPUT_PATH}
                       RUNTIME_OUTPUT_DIRECTORY ${EXECUTABLE_OUTPUT_PATH}
-                      INSTALL_NAME_DIR lib
                       OUTPUT_NAME "opencv_createsamples")
 
 if(ENABLE_SOLUTION_FOLDERS)
diff --git a/apps/traincascade/CMakeLists.txt b/apps/traincascade/CMakeLists.txt
index b21fb87367..b36d0bed60 100644
--- a/apps/traincascade/CMakeLists.txt
+++ b/apps/traincascade/CMakeLists.txt
@@ -23,7 +23,6 @@ set_target_properties(${the_target} PROPERTIES
                       DEBUG_POSTFIX "${OPENCV_DEBUG_POSTFIX}"
                       ARCHIVE_OUTPUT_DIRECTORY ${LIBRARY_OUTPUT_PATH}
                       RUNTIME_OUTPUT_DIRECTORY ${EXECUTABLE_OUTPUT_PATH}
-                      INSTALL_NAME_DIR lib
                       OUTPUT_NAME "opencv_traincascade")
 
 if(ENABLE_SOLUTION_FOLDERS)
diff --git a/cmake/OpenCVModule.cmake b/cmake/OpenCVModule.cmake
index 23a637a25a..bd3286c2a2 100644
--- a/cmake/OpenCVModule.cmake
+++ b/cmake/OpenCVModule.cmake
@@ -833,7 +833,6 @@ macro(_ocv_create_module)
     COMPILE_PDB_OUTPUT_DIRECTORY ${LIBRARY_OUTPUT_PATH}
     LIBRARY_OUTPUT_DIRECTORY ${LIBRARY_OUTPUT_PATH}
     RUNTIME_OUTPUT_DIRECTORY ${EXECUTABLE_OUTPUT_PATH}
-    INSTALL_NAME_DIR lib
   )
 
   # For dynamic link numbering convenions
diff --git a/modules/highgui/src/window_gtk.cpp b/modules/highgui/src/window_gtk.cpp
index c722165102..29172e72d1 100644
--- a/modules/highgui/src/window_gtk.cpp
+++ b/modules/highgui/src/window_gtk.cpp
@@ -1216,13 +1216,33 @@ static void icvDeleteWindow( CvWindow* window )
     }
 
     cvFree( &window );
+
+    // if last window...
+    if( hg_windows == 0 )
+    {
 #ifdef HAVE_GTHREAD
-    // if last window, send key press signal
-    // to jump out of any waiting cvWaitKey's
-    if(hg_windows==0 && thread_started){
-        g_cond_broadcast(cond_have_key);
-    }
+        if( thread_started )
+        {
+            // send key press signal to jump out of any waiting cvWaitKey's
+            g_cond_broadcast( cond_have_key );
+        }
+        else
+        {
+#endif
+            // Some GTK+ modules (like the Unity module) use GDBusConnection,
+            // which has a habit of postponing cleanup by performing it via
+            // idle sources added to the main loop. Since this was the last window,
+            // we can assume that no event processing is going to happen in the
+            // nearest future, so we should force that cleanup (by handling all pending
+            // events) while we still have the chance.
+            // This is not needed if thread_started is true, because the background
+            // thread will process events continuously.
+            while( gtk_events_pending() )
+                gtk_main_iteration();
+#ifdef HAVE_GTHREAD
+        }
 #endif
+    }
 }
 
 
diff --git a/modules/imgcodecs/src/grfmt_png.cpp b/modules/imgcodecs/src/grfmt_png.cpp
index 2f4a62baee..d5d175f89e 100644
--- a/modules/imgcodecs/src/grfmt_png.cpp
+++ b/modules/imgcodecs/src/grfmt_png.cpp
@@ -228,8 +228,6 @@ bool  PngDecoder::readData( Mat& img )
     AutoBuffer<uchar*> _buffer(m_height);
     uchar** buffer = _buffer;
     int color = img.channels() > 1;
-    uchar* data = img.ptr();
-    int step = (int)img.step;
 
     if( m_png_ptr && m_info_ptr && m_end_info && m_width && m_height )
     {
@@ -281,7 +279,7 @@ bool  PngDecoder::readData( Mat& img )
             png_read_update_info( png_ptr, info_ptr );
 
             for( y = 0; y < m_height; y++ )
-                buffer[y] = data + y*step;
+                buffer[y] = img.data + y*img.step;
 
             png_read_image( png_ptr, buffer );
             png_read_end( png_ptr, end_info );
diff --git a/modules/imgproc/src/pyramids.cpp b/modules/imgproc/src/pyramids.cpp
index 0d311e22fa..f7ff60d532 100644
--- a/modules/imgproc/src/pyramids.cpp
+++ b/modules/imgproc/src/pyramids.cpp
@@ -1010,7 +1010,7 @@ pyrUp_( const Mat& _src, Mat& _dst, int)
         for( ; sy <= y + 1; sy++ )
         {
             WT* row = buf + ((sy - sy0) % PU_SZ)*bufstep;
-            int _sy = borderInterpolate(sy*2, dsize.height, BORDER_REFLECT_101)/2;
+            int _sy = borderInterpolate(sy*2, ssize.height*2, BORDER_REFLECT_101)/2;
             const T* src = _src.ptr<T>(_sy);
 
             if( ssize.width == cn )
@@ -1031,6 +1031,11 @@ pyrUp_( const Mat& _src, Mat& _dst, int)
                 t0 = src[sx - cn] + src[sx]*7;
                 t1 = src[sx]*8;
                 row[dx] = t0; row[dx + cn] = t1;
+
+                if (dsize.width > ssize.width*2)
+                {
+                    row[(_dst.cols-1) + x] = row[dx + cn];
+                }
             }
 
             for( x = cn; x < ssize.width - cn; x++ )
@@ -1057,6 +1062,17 @@ pyrUp_( const Mat& _src, Mat& _dst, int)
             dst1[x] = t1; dst0[x] = t0;
         }
     }
+
+    if (dsize.height > ssize.height*2)
+    {
+        T* dst0 = _dst.ptr<T>(ssize.height*2-2);
+        T* dst2 = _dst.ptr<T>(ssize.height*2);
+
+        for(x = 0; x < dsize.width ; x++ )
+        {
+            dst2[x] = dst0[x];
+        }
+    }
 }
 
 typedef void (*PyrFunc)(const Mat&, Mat&, int);
diff --git a/modules/java/CMakeLists.txt b/modules/java/CMakeLists.txt
index bb3240bfdf..7b383c5491 100644
--- a/modules/java/CMakeLists.txt
+++ b/modules/java/CMakeLists.txt
@@ -386,7 +386,6 @@ set_target_properties(${the_module} PROPERTIES
     ARCHIVE_OUTPUT_DIRECTORY ${LIBRARY_OUTPUT_PATH}
     LIBRARY_OUTPUT_DIRECTORY ${LIBRARY_OUTPUT_PATH}
     RUNTIME_OUTPUT_DIRECTORY ${EXECUTABLE_OUTPUT_PATH}
-    INSTALL_NAME_DIR ${OPENCV_LIB_INSTALL_PATH}
     LINK_INTERFACE_LIBRARIES ""
     )
 
diff --git a/modules/videoio/src/cap_ffmpeg_impl.hpp b/modules/videoio/src/cap_ffmpeg_impl.hpp
index dbb7cf9ee8..8ac33d749d 100644
--- a/modules/videoio/src/cap_ffmpeg_impl.hpp
+++ b/modules/videoio/src/cap_ffmpeg_impl.hpp
@@ -1319,7 +1319,7 @@ struct CvVideoWriter_FFMPEG
     uint8_t         * picbuf;
     AVStream        * video_st;
     int               input_pix_fmt;
-    Image_FFMPEG      temp_image;
+    unsigned char   * aligned_input;
     int               frame_width, frame_height;
     int               frame_idx;
     bool              ok;
@@ -1396,7 +1396,7 @@ void CvVideoWriter_FFMPEG::init()
     picbuf = 0;
     video_st = 0;
     input_pix_fmt = 0;
-    memset(&temp_image, 0, sizeof(temp_image));
+    aligned_input = NULL;
     img_convert_ctx = 0;
     frame_width = frame_height = 0;
     frame_idx = 0;
@@ -1665,7 +1665,20 @@ static int icv_av_write_frame_FFMPEG( AVFormatContext * oc, AVStream * video_st,
 /// write a frame with FFMPEG
 bool CvVideoWriter_FFMPEG::writeFrame( const unsigned char* data, int step, int width, int height, int cn, int origin )
 {
-    bool ret = false;
+    // check parameters
+    if (input_pix_fmt == AV_PIX_FMT_BGR24) {
+        if (cn != 3) {
+            return false;
+        }
+    }
+    else if (input_pix_fmt == AV_PIX_FMT_GRAY8) {
+        if (cn != 1) {
+            return false;
+        }
+    }
+    else {
+        assert(false);
+    }
 
     if( (width & -2) != frame_width || (height & -2) != frame_height || !data )
         return false;
@@ -1679,58 +1692,29 @@ bool CvVideoWriter_FFMPEG::writeFrame( const unsigned char* data, int step, int
     AVCodecContext *c = &(video_st->codec);
 #endif
 
-#if LIBAVFORMAT_BUILD < 5231
-    // It is not needed in the latest versions of the ffmpeg
-    if( c->codec_id == CV_CODEC(CODEC_ID_RAWVIDEO) && origin != 1 )
-    {
-        if( !temp_image.data )
-        {
-            temp_image.step = (width*cn + 3) & -4;
-            temp_image.width = width;
-            temp_image.height = height;
-            temp_image.cn = cn;
-            temp_image.data = (unsigned char*)malloc(temp_image.step*temp_image.height);
-        }
-        for( int y = 0; y < height; y++ )
-            memcpy(temp_image.data + y*temp_image.step, data + (height-1-y)*step, width*cn);
-        data = temp_image.data;
-        step = temp_image.step;
-    }
-#else
-    if( width*cn != step )
+    // FFmpeg contains SIMD optimizations which can sometimes read data past
+    // the supplied input buffer. To ensure that doesn't happen, we pad the
+    // step to a multiple of 32 (that's the minimal alignment for which Valgrind
+    // doesn't raise any warnings).
+    const int STEP_ALIGNMENT = 32;
+    if( step % STEP_ALIGNMENT != 0 )
     {
-        if( !temp_image.data )
+        int aligned_step = (step + STEP_ALIGNMENT - 1) & -STEP_ALIGNMENT;
+
+        if( !aligned_input )
         {
-            temp_image.step = width*cn;
-            temp_image.width = width;
-            temp_image.height = height;
-            temp_image.cn = cn;
-            temp_image.data = (unsigned char*)malloc(temp_image.step*temp_image.height);
+            aligned_input = (unsigned char*)av_mallocz(aligned_step * height);
         }
+
         if (origin == 1)
             for( int y = 0; y < height; y++ )
-                memcpy(temp_image.data + y*temp_image.step, data + (height-1-y)*step, temp_image.step);
+                memcpy(aligned_input + y*aligned_step, data + (height-1-y)*step, step);
         else
             for( int y = 0; y < height; y++ )
-                memcpy(temp_image.data + y*temp_image.step, data + y*step, temp_image.step);
-        data = temp_image.data;
-        step = temp_image.step;
-    }
-#endif
+                memcpy(aligned_input + y*aligned_step, data + y*step, step);
 
-    // check parameters
-    if (input_pix_fmt == AV_PIX_FMT_BGR24) {
-        if (cn != 3) {
-            return false;
-        }
-    }
-    else if (input_pix_fmt == AV_PIX_FMT_GRAY8) {
-        if (cn != 1) {
-            return false;
-        }
-    }
-    else {
-        assert(false);
+        data = aligned_input;
+        step = aligned_step;
     }
 
     if ( c->pix_fmt != input_pix_fmt ) {
@@ -1738,6 +1722,7 @@ bool CvVideoWriter_FFMPEG::writeFrame( const unsigned char* data, int step, int
         // let input_picture point to the raw data buffer of 'image'
         _opencv_ffmpeg_av_image_fill_arrays(input_picture, (uint8_t *) data,
                        (AVPixelFormat)input_pix_fmt, width, height);
+        input_picture->linesize[0] = step;
 
         if( !img_convert_ctx )
         {
@@ -1762,10 +1747,11 @@ bool CvVideoWriter_FFMPEG::writeFrame( const unsigned char* data, int step, int
     else{
         _opencv_ffmpeg_av_image_fill_arrays(picture, (uint8_t *) data,
                        (AVPixelFormat)input_pix_fmt, width, height);
+        picture->linesize[0] = step;
     }
 
     picture->pts = frame_idx;
-    ret = icv_av_write_frame_FFMPEG( oc, video_st, outbuf, outbuf_size, picture) >= 0;
+    bool ret = icv_av_write_frame_FFMPEG( oc, video_st, outbuf, outbuf_size, picture) >= 0;
     frame_idx++;
 
     return ret;
@@ -1848,11 +1834,7 @@ void CvVideoWriter_FFMPEG::close()
     /* free the stream */
     avformat_free_context(oc);
 
-    if( temp_image.data )
-    {
-        free(temp_image.data);
-        temp_image.data = 0;
-    }
+    av_freep(&aligned_input);
 
     init();
 }
diff --git a/samples/data/tree.avi b/samples/data/tree.avi
index 8c7ff14a30..ded3608a51 100644
Binary files a/samples/data/tree.avi and b/samples/data/tree.avi differ