DogDetector

DogDetector is an iOS SwiftUI app that fetches dog images, runs pose detection, and renders highlighted detections with keypoints. Demo Video here: https://www.youtube.com/shorts/zKRpOKIQS7Q

Platform Support

• iOS 18+

Model

• Uses a YOLO pose estimation model trained for dogs (dog_pose_model.mlmodel).
• Dog-pose setup follows the Ultralytics Dog-Pose dataset conventions (24 keypoints, dog class).
• Reference: https://docs.ultralytics.com/datasets/pose/dog-pose/#introduction

Core Tech Stack

• SwiftUI for UI
• Apple Vision (CoreMLRequest) for model inference
• CoreML for running the YOLO pose model
• CoreImage/CoreGraphics (CGImage) for rendering overlays and effects

Detection Pipeline

1. Image enters the system as a CGImage.
2. DogDetectionService runs Vision + CoreML with .scaleFit.
3. Model output is decoded from tensor shape [1, 300, 78]:
   • per-detection box + score + keypoints
4. Coordinates are mapped from model-input space back to original image space.
5. Results are converted to normalized coordinates for drawing.

Performance Design

• Image transforms are done in CGImage space to reduce conversion overhead.
• Rendering operations (lensHighlightRegions, keypoint drawing) use CoreImage/CoreGraphics directly.
• Shared CIContext is reused for frame/image conversions.
• DogViewModel uses NSCache (count + memory cost limits) to avoid unbounded image memory growth.

UI Modes

Scroll Feed Mode

• Fetches random dog images from Dog API.
• Toggle to enable/disable detection overlays.
• Lazy list paging to load more images as user scrolls.

Live Camera Mode

• Full-screen camera preview for real-time model testing.
• Runs detection per frame through DogDetectionService.
• Applies a One Euro filter to smooth box/keypoint jitter frame-to-frame.
• Useful for validating model behavior under live framerate constraints.

Networking and Services Layer

The app uses a small service-oriented networking layer:

• Endpoint protocol defines path/method/query/body contract.
• NetworkClient builds and executes typed async requests and decodes JSON generically.
• DogService owns Dog API-specific endpoint(s) and maps response payloads to app models.
• DogDetectionService is dedicated to inference and decode logic, separated from UI/view model code.

This separation keeps API access, ML inference, and UI state management independent and easier to maintain.

Error Handling

• Detection path uses throwing async flow.
• UI surfaces service/network failures through DogViewModel.errorMessage and a SwiftUI alert.