Metal Camera Tutorial Part 1: Getting raw camera data

6 minute read

A lot of apps nowadays use iPhone and iPad cameras. Some even do pretty badass things with it (performance wise), like running each frame through a neural network or applying a realtime filter. Either way you may want to get as low as you can in terms of the level at which you interact with the device hardware, be it getting data from a camera sensor or computations involving GPU — you still want to minimise the impact on device’s limited computational resources. In this tutorial we will get raw camera frames data. If you are trying to achieve the lowest-overhead access to camera sensor, using a AVCaptureSession would be a really good start!

What is AVCaptureSession?

AVCaptureSession is a class in AVFoundation framework — whenever you write any code for audio/video recording or playback, the chances are you will end up using some parts of AVFoundation (AV standing for “Audio/Video”). As Apple reference points out, you use an AVCaptureSession object to coordinate the flow of data from audio or video input devices to outputs. So it is supposed to be a pretty generic interface between something generating media data (for example, camera sensor) and something receiving it (say, your app).

What do I do with it again?

The workflow is pretty straightforward, and basically consists of the following steps:

  • Initialise session. Session being actual AVCaptureSession instance.
  • Request access to hardware (for instance, the camera) and get a AVCaptureDevice instance. This instance represents a single hardware piece — say, a front camera or a back camera.
  • Add input to session. Input is a AVCaptureDeviceInput instance initialised with your AVCaptureDevice.
  • Add output to session. In our case it’s going to be a AVCaptureVideoDataOutput instance that you need to configure with the format you want your data in.
  • Start the session and off you go. The data should be streaming now.

Give me the code!

There is a bit more to each of those steps of course, so let’s step into more detail.

Initialise session

Well, this part is pretty easy — you simply initialise an instance.

let captureSession = AVCaptureSession()

Request access to hardware

We are still on relatively easy side of things with this one. You only need to provide the media type you want to get access to — in out case it’s AVMediaTypeVideo. This is not an enum by the way, but rather a String constant.

AVCaptureDevice.requestAccessForMediaType(AVMediaTypeVideo) {
    (granted: Bool) -> Void in
    guard granted else {
        /// Report an error. We didn't get access to hardware.
    /// All good, access granted.

Now you need to get a AVCaptureDevice instance representing the camera. I would suggest adding a routine method for that.

func device(mediaType: String, position: AVCaptureDevicePosition) -> AVCaptureDevice? {
    guard let devices = AVCaptureDevice.devices(withMediaType: mediaType) as? [AVCaptureDevice] else { return nil }
    if let index = devices.index(where: { $0.position == position }) {
        return devices[index]
    return nil

What this method does is simply gets an Array of devices of specified type and then filters out the one with requested AVCaptureDevicePosition. So now you can get a AVCaptureDevice instance representing the front camera!

guard let inputDevice = device(mediaType: AVMediaTypeVideo, position: .front) else { 
	/// Handle an error. We couldn't get hold of the requested hardware.

/// Do stuff with your inputDevice!

Add input to session

Here we initialise a AVCaptureDeviceInput instance, and try to add it to the camera session.

var captureInput: AVCaptureDeviceInput!

do {
    captureInput = try AVCaptureDeviceInput(device: inputDevice)
catch {
    /// Handle an error. Input device is not available.

Now, we are going to configure a bunch of the capture session things in the next paragraphs, and we would like it to be all applied atomically. You can do that by wrapping the configuration into beginConfiguration()/commitConfiguration() calls:


guard captureSession.canAddInput(captureInput) else {
    /// Handle an error. We failed to add an input device.


Add output to session

Now it’s time to configure the output, e.g. something that you’re going to use in your code to actually get hold of the raw camera data. It starts easy:

let outputData = AVCaptureVideoDataOutput()

This would be a good time to configure the output, for instance specify data format. In case of video streaming the big choice you have to make is the color space: you can either use RGB or YCbCr. You almost certainly heard of RGB which stands for Red Green Blue, which in its turn means we store three values for each pixel. This is actually quite a bit redundant, which is why there is another YCbCr option.

YCbCr also stores 3 components, Y being the luma component and CB and CR are the blue-difference and red-difference chroma components. Basically you get a greyscale image as Y and the color offsets in those Cb and Cr. You can get a better insight of the whole thing on Wikipedia. If you’re worried about performance and try to squeeze every bit of the processor juice you may want to use YCbCr instead of RGB — not only you get smaller textures for each frame, but also operate in hardware native format, which means it won’t use any of the device resources to convert output data to RGB.

outputData.videoSettings = [
    kCVPixelBufferPixelFormatTypeKey as AnyHashable : Int(kCVPixelFormatType_420YpCbCr8BiPlanarVideoRange)

Now it’s time to set the delegate that is going to receive each and every video frame. You will also need to specify a dispatch queue:

let captureSessionQueue = DispatchQueue(label: "CameraSessionQueue", attributes: [])
outputData.setSampleBufferDelegate(self, queue: captureSessionQueue)

You have set self as sample buffer delegate, so don’t forget to implement the AVCaptureVideoDataOutputSampleBufferDelegate protocol’s didOutputSampleBuffer method. It will be your output callback that will be called for every frame, passing the frame data in a CMSampleBuffer.

Finally, add your configured output to the session instance.

guard captureSession.canAddOutput(outputData) else {
    /// Handle an error. We failed to add an output device.


Start session

Ok, we are done with configuration for now! Let’s call commitConfiguration() and fire up the session.


Now, if you have configured everything correctly, you should start receiving sample buffers with live front camera video frames in the AVCaptureVideoDataOutputSampleBufferDelegate protocol callback.

@objc func captureOutput(_ captureOutput: AVCaptureOutput!, didOutputSampleBuffer sampleBuffer: CMSampleBuffer!, from connection: AVCaptureConnection!) {
	/// Do more fancy stuff with sampleBuffer.

That’s it? What do I do with those sample buffers?

CMSampleBuffer is a Core Foundation object representing a generic container for media data. It can contain video frames, audio or other media types, as well as context info, like timestamps, playback speed and so on. Normally you extract the data you need before working with it, and in case of video streaming this data would be frame textures.

Next time we are going to explore a way to grab those frames from a CMSampleBuffer and convert it to a texture that we can later use with Apple’s GPU-powered framework Metal.

Where do I go from here?

This was the first part of Metal Camera Tutorial series, where we explore ways of achieving lowest-overhead access to hardware to grab camera frames, convert them to textures and render on screen in real time:

You can check out the final project from this Metal Camera Tutorial on GitHub.

Follow @alexstaravoitau Star Fork Download

Leave a Comment