perception-development

To add an ONNX model to a perception analyzer pipeline, implement the OnnxRunner pattern using Python adapters and pybind11 bridges. You must delegate execution to C++ runners while enforcing strict 0-1 coordinate normalization for deterministic results.

The OnnxRunner pattern is a unified workflow for AI perception development that migrates legacy implementations into standardized pipelines. It uses Python adapters to delegate to C++ perception runners, ensuring consistent results across face, hands, and pose analyzers.

Yes, you must activate the spatial-iteration-engine conda environment to run perception analyzers. Environment discipline is strictly enforced to ensure safe, deterministic execution and proper GIL management across pybind11 bridges.

Migrating legacy analyzers to the OnnxRunner workflow involves transitioning from Pattern B implementations to the unified Pattern A architecture. This requires wrapping ONNX models in Python adapters that delegate to C++ perception runners via pybind11 bridges with GIL release.

Perception adapter failures during C++ execution often stem from improper GIL management in pybind11 bridges or incorrect coordinate normalization. Ensure the spatial-iteration-engine conda environment is active and that all perception coordinates are strictly normalized between 0 and 1.

Yes, Python adapters are used to delegate to C++ perception runners for consistent execution. This architecture relies on pybind11 bridges with GIL release to safely connect Python logic with C++ ONNX model execution in the adapters/perception directory.