iOS CVImageBuffer distorted from AVCaptureSessionDataOutput with AVCaptureSessionPresetPhoto

Question

At a high level, I created an app that lets a user point his or her iPhone camera around and see video frames that have been processed with visual effects. Additionally, the user can tap a button to take a freeze-frame of the current preview as a high-resolution photo that is saved in their iPhone library.

To do this, the app follows this procedure:

1) Create an AVCaptureSession

captureSession = [[AVCaptureSession alloc] init];
[captureSession setSessionPreset:AVCaptureSessionPreset640x480];

2) Hook up an AVCaptureDeviceInput using the back-facing camera.

videoInput = [[[AVCaptureDeviceInput alloc] initWithDevice:backFacingCamera error:&error] autorelease];
[captureSession addInput:videoInput];

3) Hook up an AVCaptureStillImageOutput to the session to be able to capture still frames at Photo resolution.

stillOutput = [[AVCaptureStillImageOutput alloc] init];
[stillOutput setOutputSettings:[NSDictionary
    dictionaryWithObject:[NSNumber numberWithInt:kCVPixelFormatType_32BGRA]
    forKey:(id)kCVPixelBufferPixelFormatTypeKey]];
[captureSession addOutput:stillOutput];

4) Hook up an AVCaptureVideoDataOutput to the session to be able to capture individual video frames (CVImageBuffers) at a lower resolution

videoOutput = [[AVCaptureVideoDataOutput alloc] init];
[videoOutput setVideoSettings:[NSDictionary dictionaryWithObject:[NSNumber numberWithInt:kCVPixelFormatType_32BGRA] forKey:(id)kCVPixelBufferPixelFormatTypeKey]];
[videoOutput setSampleBufferDelegate:self queue:dispatch_get_main_queue()];
[captureSession addOutput:videoOutput];

5) As video frames are captured, the delegate's method is called with each new frame as a CVImageBuffer:

- (void)captureOutput:(AVCaptureOutput *)captureOutput didOutputSampleBuffer:(CMSampleBufferRef)sampleBuffer fromConnection:(AVCaptureConnection *)connection
{
    CVImageBufferRef pixelBuffer = CMSampleBufferGetImageBuffer(sampleBuffer);
    [self.delegate processNewCameraFrame:pixelBuffer];
}

6) Then the delegate processes/draws them:

- (void)processNewCameraFrame:(CVImageBufferRef)cameraFrame {
    CVPixelBufferLockBaseAddress(cameraFrame, 0);
    int bufferHeight = CVPixelBufferGetHeight(cameraFrame);
    int bufferWidth = CVPixelBufferGetWidth(cameraFrame);

    glClear(GL_COLOR_BUFFER_BIT);

    glGenTextures(1, &videoFrameTexture_);
    glBindTexture(GL_TEXTURE_2D, videoFrameTexture_);
    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);

    glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA, bufferWidth, bufferHeight, 0, GL_BGRA, GL_UNSIGNED_BYTE, CVPixelBufferGetBaseAddress(cameraFrame));

    glBindBuffer(GL_ARRAY_BUFFER, [self vertexBuffer]);
    glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, [self indexBuffer]);

    glDrawElements(GL_TRIANGLE_STRIP, 4, GL_UNSIGNED_SHORT, BUFFER_OFFSET(0));

    glBindBuffer(GL_ARRAY_BUFFER, 0);
    glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, 0);
    [[self context] presentRenderbuffer:GL_RENDERBUFFER];

    glDeleteTextures(1, &videoFrameTexture_);

    CVPixelBufferUnlockBaseAddress(cameraFrame, 0);
}

This all works and leads to the correct results. I can see a video preview of 640x480 processed through OpenGL. It looks like this:

However, if I capture a still image from this session, its resolution will also be 640x480. I want it to be high resolution, so in step one I change the preset line to:

[captureSession setSessionPreset:AVCaptureSessionPresetPhoto];

This correctly captures still images at the highest resolution for the iPhone4 (2592x1936).

However, the video preview (as received by the delegate in steps 5 and 6) now looks like this:

I've confirmed that every other preset (High, medium, low, 640x480, and 1280x720) previews as intended. However, the Photo preset seems to send buffer data in a different format.

I've also confirmed that the data being sent to the buffer at the Photo preset is actually valid image data by taking the buffer and creating a UIImage out of it instead of sending it to openGL:

CGColorSpaceRef colorSpace = CGColorSpaceCreateDeviceRGB();
CGContextRef context = CGBitmapContextCreate(CVPixelBufferGetBaseAddress(cameraFrame), bufferWidth, bufferHeight, 8, bytesPerRow, colorSpace, kCGBitmapByteOrder32Little | kCGImageAlphaPremultipliedFirst); 
CGImageRef cgImage = CGBitmapContextCreateImage(context); 
UIImage *anImage = [UIImage imageWithCGImage:cgImage];

This shows an undistorted video frame.

I've done a bunch of searching and can't seem to fix it. My hunch is that it's a data format issue. That is, I believe that the buffer is being set correctly, but with a format that this line doesn't understand:

glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA, bufferWidth, bufferHeight, 0, GL_BGRA, GL_UNSIGNED_BYTE, CVPixelBufferGetBaseAddress(cameraFrame));

My hunch was that changing the external format from GL_BGRA to something else would help, but it doesn't... and through various means it looks like the buffer is actually in GL_BGRA.

Does anyone know what's going on here? Or do you have any tips on how I might go about debugging why this is happening? (What's super weird is that this happens on an iphone4 but not on an iPhone 3GS ... both running ios4.3)

Answer 1

This was a doozy.

As Lio Ben-Kereth pointed out, the padding is 48 as you can see from the debugger

(gdb) po pixelBuffer
<CVPixelBuffer 0x2934d0 width=852 height=640 bytesPerRow=3456 pixelFormat=BGRA
# => 3456 - 852 * 4 = 48

OpenGL can compensate for this, but OpenGL ES cannot (more info here openGL SubTexturing)

So here is how I'm doing it in OpenGL ES:

(CVImageBufferRef)pixelBuffer   // pixelBuffer containing the raw image data is passed in

/* ... */
glActiveTexture(GL_TEXTURE0);
glBindTexture(GL_TEXTURE_2D, videoFrameTexture_);

int frameWidth = CVPixelBufferGetWidth(pixelBuffer);
int frameHeight = CVPixelBufferGetHeight(pixelBuffer);

size_t bytesPerRow, extraBytes;

bytesPerRow = CVPixelBufferGetBytesPerRow(pixelBuffer);
extraBytes = bytesPerRow - frameWidth*4;

GLubyte *pixelBufferAddr = CVPixelBufferGetBaseAddress(pixelBuffer);

if ( [[captureSession sessionPreset] isEqualToString:@"AVCaptureSessionPresetPhoto"] )
{

    glTexImage2D( GL_TEXTURE_2D, 0, GL_RGBA, frameWidth, frameHeight, 0, GL_BGRA, GL_UNSIGNED_BYTE, NULL );

    for( int h = 0; h < frameHeight; h++ )
    {
        GLubyte *row = pixelBufferAddr + h * (frameWidth * 4 + extraBytes);
        glTexSubImage2D( GL_TEXTURE_2D, 0, 0, h, frameWidth, 1, GL_BGRA, GL_UNSIGNED_BYTE, row );
    }
}
else
{
    glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA, frameWidth, frameHeight, 0, GL_BGRA, GL_UNSIGNED_BYTE, pixelBufferAddr);
}

Before, I was using AVCaptureSessionPresetMedium and getting 30fps. In AVCaptureSessionPresetPhoto I'm getting 16fps on an iPhone 4. The looping for the sub-texture does not seem to affect the frame rate.

I'm using an iPhone 4 on iOS 5.

Answer 2

Just draw like this.

size_t bytesPerRow = CVPixelBufferGetBytesPerRow(pixelBuffer);
int frameHeight = CVPixelBufferGetHeight(pixelBuffer);

GLubyte *pixelBufferAddr = CVPixelBufferGetBaseAddress(pixelBuffer);
glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA, (GLsizei)bytesPerRow / 4, (GLsizei)frameHeight, 0, GL_BGRA, GL_UNSIGNED_BYTE, pixelBufferAddr);

Answer 3

Good point Mats. But as a matter of fact the padding is larger, it's:

bytesPerRow = 4 * bufferWidth + 48;

It works great on the iphone 4 back camera, and solved the issue sotangochips reported about.

Answer 4

Dex, thanks for the excellent answer. To make your code more generic, I would replace:

if ( [[captureSession sessionPreset] isEqualToString:@"AVCaptureSessionPresetPhoto"] )

with

if ( extraBytes > 0 )

Answer 5

I think I found your answer, and I'm sorry because it is no good news.

You can check this link: http://developer.apple.com/library/mac/#documentation/AudioVideo/Conceptual/AVFoundationPG/Articles/04_MediaCapture.html

Configuring a Session

Symbol: AVCaptureSessionPresetPhoto
Resolution: Photo.
Comments: Full photo resolution. This is not supported for video output.

Answer 6

The sessionPresetPhoto is the setting for capturing a photo with highest quality. When we use AVCaptureStillImageOutput with preset photo, the frame captured from video stream has always exactly the resolution of the iPad or iPhone screen. I have had the same problem with iPad Pro 12.9 inch which has a 2732 * 2048 resolution. That means the frame I captured from video stream was 2732 * 2048 but it was always distorted and shifted. I tried above mentioned solutions but it did not solve my problem. Finally, I realised that the width of the frame should always be divisible to 8 which 2732 is not. 2732/8 = 341.5. So what I did was to calculate the modulo of width and 8. If modulo is not equal to zero then I add it to the width. In this case 2732%8 = 4 and then I get 2732+4 = 2736. So I will set this frame width in CVPixelBufferCreate in order to initialise my pixelBuffer(CVPixelBufferRef).

Answer 7

The image buffer you get seem to contain some padding at the end. E.g.

bytesPerRow = 4 * bufferWidth + 12;

This is often done so each pixel row starts at a 16 byte offset.

Answer 8

Use this size evereywhere in your code

 int width_16 = (int)yourImage.size.width - (int)yourImage.size.width%16; 
 int height_ = (int)(yourImage.size.height/yourImage.size.width * width_16) ;
 CGSize video_size_ = CGSizeMake(width_16, height_);