How to add a CIFilter to MTLTexture Using ARMatteGenerator?

Question

I am working off of Apple's sample project related to using the ARMatteGenerator to generate a a MTLTexture that can be used as an occlusion matte in the people occlusion technology.

I would like to determine how I could run the generated matte through a CIFilter. In my code, I am "filtering" the matte like such;

func updateMatteTextures(commandBuffer: MTLCommandBuffer) {
    guard let currentFrame = session.currentFrame else {
        return
    }
    var targetImage: CIImage?
    alphaTexture = matteGenerator.generateMatte(from: currentFrame, commandBuffer: commandBuffer)
    dilatedDepthTexture = matteGenerator.generateDilatedDepth(from: currentFrame, commandBuffer: commandBuffer)
    targetImage = CIImage(mtlTexture: alphaTexture!, options: nil)
    monoAlphaCIFilter?.setValue(targetImage!, forKey: kCIInputImageKey)
    monoAlphaCIFilter?.setValue(CIColor.red, forKey: kCIInputColorKey)
    targetImage = (monoAlphaCIFilter?.outputImage)!
    let drawingBounds = CGRect(origin: .zero, size: CGSize(width: alphaTexture!.width, height: alphaTexture!.height))
    context.render(targetImage!, to: alphaTexture!, commandBuffer: commandBuffer, bounds: drawingBounds, colorSpace: CGColorSpaceCreateDeviceRGB())

}

When I go to composite the matte texture and backgrounds, there is no filtering effect applied to the matte. This is how the textures are being composited;

func compositeImagesWithEncoder(renderEncoder: MTLRenderCommandEncoder) {
    guard let textureY = capturedImageTextureY, let textureCbCr = capturedImageTextureCbCr else {
        return
    }

    // Push a debug group allowing us to identify render commands in the GPU Frame Capture tool
    renderEncoder.pushDebugGroup("CompositePass")

    // Set render command encoder state
    renderEncoder.setCullMode(.none)
    renderEncoder.setRenderPipelineState(compositePipelineState)
    renderEncoder.setDepthStencilState(compositeDepthState)

    // Setup plane vertex buffers
    renderEncoder.setVertexBuffer(imagePlaneVertexBuffer, offset: 0, index: 0)
    renderEncoder.setVertexBuffer(scenePlaneVertexBuffer, offset: 0, index: 1)

    // Setup textures for the composite fragment shader
    renderEncoder.setFragmentBuffer(sharedUniformBuffer, offset: sharedUniformBufferOffset, index: Int(kBufferIndexSharedUniforms.rawValue))
    renderEncoder.setFragmentTexture(CVMetalTextureGetTexture(textureY), index: 0)
    renderEncoder.setFragmentTexture(CVMetalTextureGetTexture(textureCbCr), index: 1)
    renderEncoder.setFragmentTexture(sceneColorTexture, index: 2)
    renderEncoder.setFragmentTexture(sceneDepthTexture, index: 3)
    renderEncoder.setFragmentTexture(alphaTexture, index: 4)
    renderEncoder.setFragmentTexture(dilatedDepthTexture, index: 5)

    // Draw final quad to display
    renderEncoder.drawPrimitives(type: .triangleStrip, vertexStart: 0, vertexCount: 4)
    renderEncoder.popDebugGroup()
}

How could I apply the CIFilter to only the alphaTexture generated by the ARMatteGenerator?

Answer 1

I don't think you want to apply a CIFilter to the alphaTexture. I assume you're using Apple's Effecting People Occlusion in Custom Renderers sample code. If you watch this year's Bringing People into AR WWDC session, they talk about generating a segmentation matte using ARMatteGenerator, which is what is being done with alphaTexture = matteGenerator.generateMatte(from: currentFrame, commandBuffer: commandBuffer). alphaTexture is a MTLTexture that is essentially an alpha mask for where humans have been detected in the camera frame (i.e. complete opaque where a human is and completely transparent where a human is not).

Adding a filter to the alpha texture won't filter the final rendered image but will simply affect the mask that is used in the compositing. If you're trying to achieve the video linked in your previous question, I would recommend adjusting the metal shader where the compositing occurs. In the session, they point out that they compare the dilatedDepth and the renderedDepth to see if they should draw virtual content or pixels from the camera:

fragment half4 customComposition(...) {
    half4 camera = cameraTexture.sample(s, in.uv);
    half4 rendered = renderedTexture.sample(s, in.uv);
    float renderedDepth = renderedDepthTexture.sample(s, in.uv);
    half4 scene = mix(rendered, camera, rendered.a);
    half matte = matteTexture.sample(s, in.uv);
    float dilatedDepth = dilatedDepthTexture.sample(s, in.uv);

    if (dilatedDepth < renderedDepth) { // People in front of rendered
        // mix together the virtual content and camera feed based on the alpha provided by the matte
        return mix(scene, camera, matte);
    } else {
        // People are not in front so just return the scene
        return scene
    }
}

Unfortunately, this is done sightly differently in the sample code, but it's still fairly easy to modify. Open up Shaders.metal. Find the compositeImageFragmentShader function. Toward the end of the function you'll see half4 occluderResult = mix(sceneColor, cameraColor, alpha); This is essentially the same operation as mix(scene, camera, matte); that we saw above. We're deciding if we should use a pixel from the scene or a pixel from camera feed based on the segmentation matte. We can easily replace the camera image pixel with an arbitrary rgba value by replacing cameraColor with a half4 that represents a color. For example, we could use half4(float4(0.0, 0.0, 1.0, 1.0)) to paint all of the pixels within the segmentation matte blue:

…
// Replacing camera color with blue
half4 occluderResult = mix(sceneColor, half4(float4(0.0, 0.0, 1.0, 1.0)), alpha);
half4 mattingResult = mix(sceneColor, occluderResult, showOccluder);
return mattingResult;

Of course, you can apply other effects as well. Dynamic grayscale static is pretty easy to achieve.

Above compositeImageFragmentShader add:

float random(float offset, float2 tex_coord, float time) {
    // pick two numbers that are unlikely to repeat
    float2 non_repeating = float2(12.9898 * time, 78.233 * time);

    // multiply our texture coordinates by the non-repeating numbers, then add them together
    float sum = dot(tex_coord, non_repeating);

    // calculate the sine of our sum to get a range between -1 and 1
    float sine = sin(sum);

    // multiply the sine by a big, non-repeating number so that even a small change will result in a big color jump
    float huge_number = sine * 43758.5453 * offset;

    // get just the numbers after the decimal point
    float fraction = fract(huge_number);

    // send the result back to the caller
    return fraction;
}

(taken from @twostraws ShaderKit)

Then modify compositeImageFragmentShader to:

…
float randFloat = random(1.0, cameraTexCoord, rgb[0]);

half4 occluderResult = mix(sceneColor, half4(float4(randFloat, randFloat, randFloat, 1.0)), alpha);
half4 mattingResult = mix(sceneColor, occluderResult, showOccluder);
return mattingResult;

You should get:

Finally, the debugger seems to have a hard time keeping up with the app. For me, when running attached Xcode, the app would freeze shortly after launch, but was typically smooth when running on its own.