Real-time camera viewfinder filters in Native code
This article explains how to create real-time camera filters for Windows Phone 8, using native code (C++).
Windows Phone 8
Note: This is an "internal" entry in the Windows Phone 8 Wiki Competition 2012Q4. The author is a Nokia / Microsoft employee.
One of the big new feature of Windows Phone 8 SDK is the support for C and C++, also known as native code support. In this article, we will have a brief look how one can leverage that support to create real-time filters for the camera. For sake of simplicity, the example will implement a simple gray filter, that will convert on the fly the camera input and show the result on the phone display.
The demo application will look like this:
Why Native filters?
Microsoft has published a very similar example, where they do live conversion of the camera viewfinder images to grayscale. The example was written for Windows Phone 7, but it also works well in WP8. Get it here. This works well, but the gray filter is quite simple, more complicated filters will require more computation, and the CPU is quickly maxed out when we try to process the camera input at several frames per seconds. The speed gain by going closer to the metal may be needed for complex algorithm. Also, you might already have your own image filters written in C and/or C++ for other platforms, that you can reuse without converting them to C#. Finally, as we will see in other wiki entries, the native side opens further optimization possibilities, like using DirectX or the ARM Neon instruction set.
Setting up the viewfinder
Our UI will be XAML based. The UI will control everything, while the C++ side is rather dumb, simply executing the filtering when asked to. Let's first create the projects for both the XAML and the C++ components:
- Start by creating a new project, of type Windows Phone App. You will find the template under the Visual C#/Windows Phone category. That's where our UI will be coded.
- Add a new project to your solution, of type Windows Phone Runtime Component. That template is under the Visual C++/Windows Phone category. That will become our image filter.
We then add a live camera stream to our UI. That is easily done by :
- In your XAML, define a rectangle that will be painted using a video brush:
<Grid x:Name="LayoutRoot" Background="Transparent">
<Rectangle Width="640" Height="480" Canvas.ZIndex="1">
<VideoBrush x:Name="viewfinderBrush" />
- In the page loaded event, Create a PhotoCaptureDevice, and set it as the source of the video brush:
Windows.Foundation.Size resolution = new Windows.Foundation.Size(640, 480);
m_camera = await PhotoCaptureDevice.OpenAsync(CameraSensorLocation.Back, resolution);
By now, with these 10 lines of codes, you should have an application with a functional camera ! Note that in the last step, we use the Windows PRT class Windows.Phone.Media.Capture.PhotoCaptureDevice which is new to Windows Phone 8. In WP7, one would have to use the Silverlight/.NET class Microsoft.Devices.PhotoCamera. Because it's a Win PRT class, it can be accessed by both managed and native code, we will soon take advantage of that possibility.
Displaying the filtered frames
The videobrush is easy to use but it doesn't offer us a way to get in-between the camera and the brush, to apply our filter. We need to find another way to display the modified frames coming from the camera. Several strategies are possible, but the strategy chosen must have a reasonable camera lag (the time between the photons enters the camera until the scene is displayed on the screen) as well as a decent frame rate. Let's have a quick look at 3 possible strategies:
- Using an Image Control, with WriteableBitmap as a source
- Using a MediaElement Control, with a custom MediaStreamSource as a source.
- Using a DirectX texture
Using an Image Control
This is the strategy used in the MSDN example. The UI control defined in XAML is :
<Image x:Name="MyCameraViewfiner" Width="640" Height="480">
The source of the Image is set to a WriteableBitmap. When the frames from the camera have been filtered, the resulting array of pixels is copied into to that WriteableBitmap.
Deployment.Current.Dispatcher.BeginInvoke(delegate() // Switch to UI thread
// Copy to WriteableBitmap.
The problem with this is the switch to UI thread (Deployment.Current.Dispatcher.BeginInvoke). The switch is required so that the application does not update the bitmap while the UI thread is doing something with it. Switching between threads is quite slow, and the lag associated with that solution was too high for the application.
Using a MediaElement
The MediaElement is 'the' UI control that is used for media playback, videos or audio, streamed media or from a local file. It takes care of all the buffering, audio sync and displaying logic and it's highly optimized. By defining a custom MediaStreamSource, we can feed the filtered camera frames to the MediaElement that will take care of displaying them properly.
Margin="4" Width="640" Height="480" />
Defining your own MediaStreamSource requires a little bit of work, but it's all worth it. The lag with this solution is small, and the frame rate we can achieve is pretty good. This is the solution we will take into use.
Using a DirectX texture
When it comes to camera lag and performance, this is probably the best solution. However DirectX has a fairly steep learning curve, and to keep things simple we will not use it. Look over this article if you're interested of going DirectX.
Our own MediaStreamSource
To feed video frames to the MediaElement we need to define our custom MediaStreamSource. Pete Brown walks us through how this should be done in his blog. Check it out for more details, here we will only go through the big lines.
Our MediaStreamSource will define a single stream of type video (RGBA). Since our real source for video data is the camera, we have a source of infinite length and that can't be seeked. Our initialization will look like this:
mediaStreamAttributes[MediaStreamAttributeKeys.VideoFourCC] = "RGBA";
mediaStreamAttributes[MediaStreamAttributeKeys.Width] = _frameWidth.ToString();
mediaStreamAttributes[MediaStreamAttributeKeys.Height] = _frameHeight.ToString();
// a zero timespan is an infinite video
// Can't seek.
mediaSourceAttributes[MediaSourceAttributesKeys.CanSeek] = false.ToString();
Whenever the MediaElement decides it needs a new video frame, it will call the GetSampleAsync() method from our MediaStreamSource.
We will implement the following sequence diagram :
To get the data of the frames coming from the camera, we will have to call to the platform functionality Windows::Phone::Media::Capture::ICameraCaptureDevice::GetPreviewBufferArgb. Let's look at its definition:
void GetPreviewBufferArgb(Platform::WriteOnlyArray<int, 1U>^ pixels)
That method fills an array of our choice with the camera data. That's a copy operation we can't avoid. The buffer is of type Platform::WriteOnlyArray, which according to MSDN documentation, is to be used when the caller passes an array for the method to fill . Makes sense. We will also use that buffer type to communicate between our managed component and our native component
The public interface of our native component will be:
public ref class WindowsPhoneRuntimeComponent sealed
void Initialize(Windows::Phone::Media::Capture::PhotoCaptureDevice^ captureDevice);
void NewViewfinderFrame( Platform::WriteOnlyArray<int,1U>^ frameData);
On the managed side, we allocate the buffer when the application is initialized, and reuse that buffer for every frames:
m_frameData = new int[(int)m_camera.PreviewResolution.Height * (int)m_camera.PreviewResolution.Width];
Every time the camera subsystem fires a PreviewFrameAvailable event we will call the native NewViewfinderFrame method who will take care of filling the buffer with camera data and filter that data.
When the UI side receives back the filtered camera data buffer, it is ready to be displayed. The UI side will do that with the following :
m_processingFrame = false;
Again, an unavoidable copy! This time the data is copied to a WriteableBitmap, m_wb. At the initialization phase, we defined m_wb as the source for the XAML Image component that displays our filtered viewfinder.
So, we end up copying two times the camera data, first from camera subsystem into our application, then from our application to the display subsystem.
Note that for this example, my data buffer is in the RGBA format. It's a format easy to handle, and the most familiar for most of us. However, it is not very efficient in terms of size and image manipulation performance. Using YUV format/color space would make more sense.
That covers the big lines of the data handling of the application. I skipped the not-so-interesting code, get the source code of the full project from the link in top right corner of this page. The last thing to do is the filtering itself.
The filtering in C++
For the filtering, I took the code from Nils Pipenbrinck's excellent blog entry on Neon optimization.
void WindowsPhoneRuntimeComponent::ConvertToGrayOriginal( Platform::WriteOnlyArray<int,1U>^ frameData)
uint8 * src = (uint8 *) frameData->Data;
uint8 * dest = (uint8 *) frameData->Data;
int n = frameData->Length;
for (i=0; i<n; i++)
int r = *src++; // load red
int g = *src++; // load green
int b = *src++; // load blue
// build weighted average:
int y = (r*77)+(g*151)+(b*28);
// undo the scale by 256 and write to memory:
*dest++ = (y>>8);
*dest++ = (y>>8);
*dest++ = (y>>8);
If you look in the project code, you will find the same filter, but optimized with the ARM Neon instruction set. I'll let you pick the one you prefer.
Wrapping it up
Hopefully this short article will help you getting started writing native filters with Windows Phone 8. Remember to always keep the performance in mind, these types of applications are very CPU intensive. Using C++ for your filters is one way to improve the performance. If the performance gain is not enough for your use cases, you might have a look at hooking your camera directly into DirectX, or optimizing your filter with Neon.
The source code is maintained in Nokia Project. Look for the zip file at the bottom of the page.