Ios App

Important: Export with nms=False — the NMS head doesn't convert cleanly. Implement NMS in Swift (~30 lines).

INT8 Quantization (halves size, ~same accuracy)

import coremltools as ct
import coremltools.optimize.coreml as cto

mlmodel = ct.models.MLModel("yolov8s_volleyball_v2.mlpackage")
config = cto.OptimizationConfig(
    global_config=cto.OpLinearQuantizerConfig(mode="linear_symmetric")
)
mlmodel_q = cto.linear_quantize_weights(mlmodel, config)
mlmodel_q.save("YOLOVolleyball_INT8.mlpackage")

Convert TrackNetV2

Use existing src/training/onnx_export.py to get ONNX, then:

mlmodel = ct.convert("tracknetv2.onnx",
    inputs=[ct.ImageType(name="input", shape=(1, 9, 288, 512))],
    minimum_deployment_target=ct.target.iOS16)
mlmodel.save("TrackNetV2.mlpackage")

Expected Performance

Critical: Drop imgsz from 1280 to 640 for mobile. 1280 is ~80ms/frame (below real-time).

Phase 2: Swift Core (4-6 weeks)

Key Components to Port

Camera + CoreML Pipeline

let model = try VNCoreMLModel(for: YOLOVolleyball().model)
let request = VNCoreMLRequest(model: model) { request, error in
    // Parse detections, run through Kalman filter
}
request.imageCropAndScaleOption = .scaleFit

// AVCaptureVideoDataOutputSampleBufferDelegate:
func captureOutput(_ output: AVCaptureOutput,
                   didOutput sampleBuffer: CMSampleBuffer, ...) {
    let pixelBuffer = CMSampleBufferGetImageBuffer(sampleBuffer)!
    let handler = VNImageRequestHandler(cvPixelBuffer: pixelBuffer)
    try? handler.perform([request])
}

Pose Tracking (replaces MediaPipe entirely)

Apple's VNDetectHumanBodyPoseRequest provides all 17 landmarks used in pose_tracker.py (shoulders, elbows, wrists, hips, knees, ankles) with zero additional models.

What to Drop

• byte_track.py — redundant for single-ball tracking, Kalman filter handles it
• cv2.createBackgroundSubtractorMOG2 — use frame differencing via vImage instead
• TTA (augment=True) — not available in CoreML, compensate with lower confidence threshold

Phase 3: UI + UX (3-4 weeks)

• SwiftUI for all screens (far less code than UIKit)
• Live camera overlay: ball trail, pose skeleton, speed readout
• Video import from Photos library for post-game analysis
• Session recording and results export
• Settings screen mapping to config.yaml thresholds

Phase 4: TestFlight + App Store (2-3 weeks)

Required Entitlements

• NSCameraUsageDescription — camera access
• NSPhotoLibraryUsageDescription — if processing saved videos
• PrivacyInfo.xcprivacy — required since iOS 17

App Size Budget

• YOLOv8s INT8: ~13MB mlpackage
• YOLOv8m INT8: ~26MB mlpackage
• TrackNetV2: ~8MB
• Total well under 200MB OTA limit

Gotchas

• CoreML warm-up takes 200-800ms on first inference — show loading state
• No Python interpreter allowed (Guideline 2.5.2)
• Test on physical devices — Simulator can't test Neural Engine or camera
• Coordinate space transforms between Vision (0-1 normalized), AVFoundation (pixel), and detection output need careful handling

Total Timeline

11-16 weeks solo with iOS experience. Add 4-6 weeks if Swift is new.