vision-coreml-recognition-workflow

Vision Core ML Recognition Workflow

Purpose

Guide custom image-model integration through Vision while keeping Core ML model execution, Vision image semantics, preprocessing, postprocessing, evaluation, and capture ownership explicit.

When To Use

• Use for custom Core ML image classification, object detection, segmentation, feature outputs, model loading, crop-and-scale, compute, evaluation, or recognition repair.
• Recommend vision-image-analysis-workflow when an Apple-provided Vision request already owns the analysis.
• Recommend direct Core ML guidance only when the input is not image-oriented or Vision does not support the model contract.

Single-Path Workflow

1. Classify the model output:
   • classification labels
   • detected objects and boxes
   • semantic or instance segmentation
   • image-to-image or pixel-buffer output
   • feature value or embedding
   • multi-output or model-specific postprocessing
2. Apply the Apple docs gate:
   • read current Vision and Core ML documentation for the model and platform
   • prefer CoreMLRequest and the current Swift Vision API for new code when compatible
   • preserve VNCoreMLModel and VNCoreMLRequest as explicit original-API repair inputs
   • inspect MLModelDescription, image constraints, metadata, and MLModelConfiguration
   • apply ../../shared/references/apple-vision-analysis-contract.md
3. Establish provenance and constraints:
   • establish immutable provenance for every shipped or downloaded model
   • record model source, version, license, checksum or immutable revision, labels, expected color space, dimensions, flexible constraints, output meanings, and known evaluation limits
   • pin the shipped model and document any compilation or download boundary
4. Configure execution:
   • select compute units from actual compatibility, energy, latency, memory, and profiling requirements
   • choose crop-and-scale behavior deliberately and preserve its inverse for output coordinates
   • keep typed Vision observations or Core ML feature values until the consumer boundary
5. Interpret and evaluate:
   • define thresholds and model-specific postprocessing from validation evidence
   • separate classification confidence from calibrated probability
   • map detection boxes or masks through preprocessing and display transforms
   • run representative fixtures and a small regression or evaluation sanity check whenever model or request logic changes
6. Return documented behavior, provenance, model contract, request family, preprocessing and postprocessing, compute plan, evaluation evidence, performance findings, and handoffs.