Real-Time AI Video Segmentation for Mobile Apps

Real-Time AI Video Segmentation for Mobile Apps

Imagine a video call with a blurred background, but instead of a smooth picture—stuttering and artifacts. Or an augmented reality app that can't keep up with user movements. These are typical consequences of poorly optimized AI segmentation. We often receive requests for real-time video segmentation—when the app understands what's in the frame: person, background, car, road, and does it for every frame at 15–30 FPS. Making it "work on a demo" is easy. Making it "doesn't overheat, doesn't lag, works on iPhone XR" requires serious optimization. Choosing the right model and a smart processing pipeline are key to a successful project.

How to Choose a Segmentation Model?

Semantic segmentation—each pixel is assigned a class (background, person, car). Applications: background replacement in video calls, AR effects, road scene analysis. Instance segmentation—a separate mask for each object of the same class (three cars—three masks). Applications: object counting, tracking. Panoptic segmentation—a combination. Heavier, rarely used on mobile.

YOLOv8n-seg is 2x better in quality than MobileNetV3 at similar speed. SAM is only for interactive segmentation, not real-time.

Why Is Performance Critical?

Mobile GPUs are limited in heat dissipation and power consumption. A naive implementation quickly leads to overheating and FPS drops. Optimization starts with model selection: lightweight architectures (MobileNetV3, YOLOv8n-seg) give 20–30 FPS on modern chips, but require a smart pipeline. According to Core ML documentation, proper Metal render configuration reduces CPU load by 40%.

How We Optimize the iOS Pipeline

class RealtimeSegmentationProcessor {

    private let model: VNCoreMLModel
    private let processQueue = DispatchQueue(label: "segmentation.process", qos: .userInteractive)

    // Frame skipping: process every N-th frame
    private var frameCounter = 0
    private let processEveryNFrames = 2  // 30fps camera → 15fps processing

    func captureOutput(_ output: AVCaptureOutput,
                       didOutput sampleBuffer: CMSampleBuffer,
                       from connection: AVCaptureConnection) {
        frameCounter += 1
        guard frameCounter % processEveryNFrames == 0 else { return }
        guard let pixelBuffer = CMSampleBufferGetImageBuffer(sampleBuffer) else { return }

        processQueue.async { [weak self] in
            self?.runSegmentation(on: pixelBuffer)
        }
    }

    private func runSegmentation(on pixelBuffer: CVPixelBuffer) {
        let request = VNCoreMLRequest(model: model) { [weak self] req, _ in
            guard let observation = req.results?.first as? VNCoreMLFeatureValueObservation,
                  let maskArray = observation.featureValue.multiArrayValue else { return }

            let mask = self?.processMask(maskArray)
            DispatchQueue.main.async {
                self?.delegate?.didUpdateSegmentationMask(mask)
            }
        }

        // Important: pixelBuffer must be in the correct format
        request.imageCropAndScaleOption = .scaleFill
        let handler = VNImageRequestHandler(cvPixelBuffer: pixelBuffer,
                                           orientation: .right)  // landscape orientation
        try? handler.perform([request])
    }

    private func processMask(_ array: MLMultiArray) -> SegmentationMask {
        // Convert MLMultiArray → CVPixelBuffer for rendering
        // Shape: [numClasses, height, width]
        let numClasses = array.shape[0].intValue
        let height = array.shape[1].intValue
        let width = array.shape[2].intValue

        // Argmax across classes for each pixel → label map
        var labelMap = [UInt8](repeating: 0, count: height * width)
        for y in 0..<height {
            for x in 0..<width {
                var maxClass = 0
                var maxVal: Float = -Float.infinity
                for c in 0..<numClasses {
                    let val = array[[c, y, x] as [NSNumber]].floatValue
                    if val > maxVal { maxVal = val; maxClass = c }
                }
                labelMap[y * width + x] = UInt8(maxClass)
            }
        }
        return SegmentationMask(labels: labelMap, width: width, height: height,
                                classColors: Self.classColorMap)
    }
}

What Is Metal Mask Rendering?

Naive approach—drawing the mask in a CPU loop—gives 3–5 FPS on rendering. The right way is Metal / OpenGL ES. A Metal shader applies the mask on the GPU without CPU involvement, ensuring stable 30 FPS even on iPhone 11.

// Metal shader for overlaying mask on video
// Inputs: videoTexture (YCbCr), maskTexture (label map), colorLUT (class→color)
fragment float4 segmentationOverlay(
    VertexOut in [[stage_in]],
    texture2d<float> videoTexture [[texture(0)]],
    texture2d<uint> maskTexture [[texture(1)]],
    texture1d<float> colorLUT [[texture(2)]],
    constant OverlayParams& params [[buffer(0)]]
) {
    float2 uv = in.texCoords;
    float4 videoColor = videoTexture.sample(sampler, uv);
    uint classLabel = maskTexture.sample(nearestSampler, uv).r;

    if (classLabel == 0) { return videoColor; }  // background—unchanged

    float4 maskColor = colorLUT.sample(sampler, float(classLabel) / float(params.numClasses));
    return mix(videoColor, maskColor, params.overlayAlpha);  // blending
}

Background Replacement on Android: MediaPipe

For video calls, binary segmentation (person/background) is popular. We use MediaPipe Selfie Segmentation—a ready-made solution optimized for this task:

// Android: MediaPipe Selfie Segmentation
val options = ImageSegmenterOptions.builder()
    .setBaseOptions(BaseOptions.builder()
        .setModelAssetPath("selfie_segmentation.tflite")
        .setDelegate(Delegate.GPU)
        .build())
    .setRunningMode(RunningMode.LIVE_STREAM)
    .setResultListener { result, _ ->
        val confidenceMask = result.confidenceMasks?.get(0)
        updateBackground(confidenceMask)
    }
    .build()

val segmenter = ImageSegmenter.createFromOptions(context, options)

Delegate.GPU is critical: on CPU, the same MediaPipe delivers 8–12 FPS; on GPU, 25–30 FPS.

Workflow: From Idea to Production

1. Task analysis—define object classes, target devices, FPS requirements.
2. Model selection—test 3–5 models on real devices, benchmark speed and mIoU.
3. Pipeline optimization—configure frame skipping, Metal rendering, TFLite GPU delegate.
4. Integration—embed into the app, error handling (graceful fallback to CPU under throttling).
5. QA and deployment—stress testing, App Store / Google Play release.

What's Included

• Model selection and calibration for your data.
• Implementation of the capture and frame processing pipeline.
• GPU mask rendering (Metal on iOS, OpenGL/Vulkan on Android).
• Integration documentation and support.
• 30-day post-release support.

Timelines and Cost

Basic single-class segmentation (e.g., person) with a ready-made model—from 1 week. Custom multi-class segmentation with Metal/GPU rendering and iOS + Android support—from 2 to 4 weeks. Cost is calculated individually after evaluating your project.

We have 5 years of mobile development experience and over 50 AI-powered projects. We guarantee stable 30 FPS on devices not older than iPhone XR and Samsung S10. Contact us for a technical consultation and accurate estimate. Request a pilot project to evaluate quality on real data.

Machine Learning in Mobile Apps: CoreML, TFLite, and On-Device Models

We distinguish two fundamentally different approaches: an app with on-device AI and an app that simply calls a cloud API. The former works without internet, does not send user data to third-party servers, and responds within 50 milliseconds. The latter depends on network latency and pricing plans. Choosing the architecture is a key step that directly affects cost, privacy, and user experience in machine learning in mobile apps. Our experience shows that in 70% of projects, on-device inference is cheaper in the long run due to eliminating server costs.

How to Choose Between CoreML and TFLite for On-Device Inference?

CoreML — Apple's native framework for running ML models on device. Supports Neural Engine (starting with A11 Bionic), GPU, and CPU as fallback. Models are converted to .mlmodel format via coremltools from PyTorch, ONNX, or TensorFlow. Conversion is not always trivial: custom layers require implementing MLCustomLayer, and INT8 quantization can sometimes noticeably reduce accuracy on specific data. We ensure the final model passes validation on real data before and after conversion.

TensorFlow Lite — cross-platform alternative for Android and Flutter. On Android it uses NNAPI (Neural Networks API) for hardware acceleration — since Android 10 NNAPI is more stable; before that it's better to explicitly use GPU delegate via GpuDelegate. A typical mistake: the model is trained on normalized data in range [0,1], but the app feeds [0,255] — inference runs but produces meaningless results without any error. We include an automatic input data validation module in the SDK.

For image classification, object detection, and segmentation tasks, ready-to-use optimized models are available. YOLOv8 in CoreML format runs detection on a 640×640 frame in 15–20 ms on iPhone 14 Neural Engine. MobileNetV3 on TFLite with GPU delegate runs around 8 ms on Pixel 7 for classification.

What If You Need Text Generation On-Device?

Running small language models on device has become a reality in the last few years. Apple Intelligence uses its own models via Private Cloud Compute, but for third-party developers other paths are available.

llama.cpp with Metal backend on iOS is a working approach for phi-3-mini (3.8B parameters, 4-bit quantization, ~2.3 GB). Inference: 15–25 tokens/second on iPhone 15 Pro. For integration in Swift, use the Swift Package llama.swift or a wrapper via C interface llama.h. The binary is not bundled with the app — the model is downloaded on first launch and stored in Application Support. Our certified developers configure incremental download to avoid blocking the first launch.

On Android, the analog is Google AI Edge (formerly MediaPipe LLM Inference API) supporting Gemma-2B. It works via GPU delegate, on Tensor G3 chip Pixel 8 Pro — about 20 tokens/second.

Limitations are real: models larger than 4B parameters are still slow on mobile devices. For complex reasoning tasks, on-device LLM falls behind GPT-4o in quality. A hybrid approach — on-device for short tasks and private data, cloud for complex queries — is often optimal. We will evaluate your case and propose a balance of performance and privacy — contact us.

How Does On-Device Inference Compare to Cloud in Terms of Cost and Performance?

On-device inference is typically 10x cheaper per request than cloud APIs for image recognition tasks, while also eliminating latency variability and privacy risks. The table below summarizes the trade-offs.

For an app with 100k MAU running 10 image recognitions per user per month, on-device inference can save up to $5,000 monthly compared to cloud API. Get a free consultation on your ML architecture today.

Integrating OpenAI API and Other Cloud Models

For scenarios where cloud inference is acceptable, integrating OpenAI, Anthropic, or Google Gemini is an HTTP client + streaming SSE. In Swift, AsyncThrowingStream is convenient for streaming responses. In Kotlin, use Flow.

Critically: API keys must never be stored in the app bundle. Even an obfuscated key can be extracted from the IPA in 10 minutes using strings or frida. Correct architecture: mobile app → your own backend → OpenAI API. The backend controls rate limiting, logs requests, and protects the key.

What Is Included in the Work (Deliverables)

• Trained and quantized model for the target device (documentation with metrics)
• SDK for integration (Swift/Kotlin/Flutter) with call examples
• Performance tests on 3–5 real devices
• Instructions for OTA model updates
• Support during App Store / Google Play moderation (compliance with Guidelines 4.2, 5.1)
• 2 weeks of technical support after release

Typical Project Pipeline

1. Task analysis — measure latency, privacy, size, supported devices.
2. Model prototyping — in Python, evaluate accuracy on target data.
3. Conversion and quantization — for CoreML/TFLite with validation.
4. Integration into the app — model wrapped in a service layer (easy to swap CoreML ↔ TFLite ↔ cloud).
5. Testing — on real devices, measure FPS, RAM, battery.
6. Deployment — via TestFlight / Firebase App Distribution, monitor metrics.

Timelines: integration of a ready CoreML/TFLite model — 1–2 weeks, development of a custom model with mobile optimization — from 6 weeks, on-device LLM chat with personalization — 4–8 weeks.

Why We Take on Complex Cases?

10+ years of experience in mobile development, 50+ implemented AI/ML solutions, guarantee of compatibility with current iOS and Android versions. All projects undergo code review and load testing. The cost includes preparation of moderation documentation and training of your team.

Contact us — we will help you choose the architecture and implement ML in your app turnkey. Order an audit of your existing solution — we will assess the potential for server cost savings free of charge. In some projects, savings can reach significant amounts per month.