axiom-vision

What it does

Guides you through implementing computer vision tasks using Vision APIs, including subject segmentation with instance masks, isolating objects from hands, hand/body pose detection, multi-person segmentation, text recognition (OCR), barcode detection, and document scanning. Also covers when to use VisionKit versus the Vision framework, and how to combine with CoreImage for compositing and HDR-like workflows. It presents several usage patterns (subject lifting, programmatic access, per-subject masks, handling multiple people, and gesture-based triggers) and decision trees to select appropriate APIs for given goals.

How it works

• Recommends the appropriate API paths based on the target task (e.g., VNGenerateForegroundInstanceMaskRequest for subject segmentation, VNDetectHumanHandPoseRequest for hand gestures, VNDetectHumanBodyPoseRequest for body pose, VNRecognizeTextRequest for OCR, VNDetectBarcodesRequest for barcodes, RecognizeDocumentsRequest for structured document data).
• Advises on setup and threading: never run Vision on the main thread; use a background queue with a VNImageRequestHandler and perform specific requests, then dispatch UI updates on the main thread.
• Provides concrete usage patterns and code snippets for:
  • Isolating an object from a hand using a subject mask and hand pose to create an exclusion mask.
  • VisionKit-based subject lifting with UI interactions.
  • Programmatic subject access via ImageAnalyzer and analysis.subjects.
  • Using VNGenerateForegroundInstanceMaskRequest for custom compositing with CoreImage (HDR-like workflows).
  • Tap-to-select and per-subject masks for multi-subject scenes.
  • Hand gesture detection (pinch) for custom triggers.
  • Separate masks for multiple people (up to 4) and fallback for crowded scenes.
  • Body pose-based action classification workflows and preparing data for a CreateML model.
• Contains decision trees and checklists to select APIs depending on needs (e.g., subject isolation, text recognition, document data extraction, whether to use VisionKit or Vision).

When to use it

• Isolate subjects from backgrounds (subject lifting) and exclude hands from object boxes.
• Detect and track hand poses or body poses for gestures or activity classification.
• Segment multiple people separately or apply per-person effects.
• Recognize text in images (OCR) or detect barcodes/QR codes.
• Scan documents with perspective correction and extract structured data (iOS 26+).
• Build live scanning experiences with DataScannerViewController or choose between VisionKit and Vision for processing.
• Combine Vision with CoreImage for compositing or HDR-like visual effects.

What it can touch

• Tools and APIs referenced include: VNGenerateForegroundInstanceMaskRequest, VNImageRequestHandler, VNDetectHumanHandPoseRequest, VNRecognizeTextRequest, VNDetectBarcodesRequest, DataScannerViewController, RecognizeDocumentsRequest, VNDocumentCameraViewController, VNGeneratePersonInstanceMaskRequest, VNDetectHumanBodyPoseRequest, ImageAnalyzer, ImageAnalysisInteraction, ImageAnalyzer.Configuration, CreateML preparation (for pose-based models).
• No file system or external services are specified beyond these VisionKit/Vision APIs; the material shows code snippets and usage patterns rather than file-level actions.

Caveats

• Platform/version requirements are specified for each API (e.g., iOS 14+, iOS 17+ for subject segmentation, iOS 26+ for RecognizeDocumentsRequest).
• Emphasizes not running Vision work on the main thread and the need to handle confidence checks and coordinate systems when converting between formats.
• Includes risk-labeled patterns (e.g., avoid training custom models when Vision APIs exist, avoid processing every frame, avoid enabling all barcode symbologies when unnecessary).
• License stated as MIT; compatibility notes list the supported OS versions and APIs.