Open-source visual quality inspection with entropy-gated kNN refinement

A contributor to the open-source hmg-physical-quality-decision-ai project released a prototype that detects objects as oriented bounding boxes and classifies them GOOD or BAD, starting with weld inspection, and is seeking feedback on its evaluation. The decision pipeline aggregates predictions across test-time augmentations, uses prediction entropy to decide when to consult a reference embedding bank, and applies kNN voting to refine uncertain predictions. The contributor asked how to compare classifier-only, always-on kNN, and entropy-gated kNN on the same detected crops, and how to measure rare-defect recall while separating detection misses from classification errors.

Hi everyone I’m a contributor to hmg-physical-quality-decision-ai, sharing this in a personal capacity. The project detects objects as oriented bounding boxes and classifies them as GOOD or BAD, with weld inspection as the first use case. It includes a CLI, API, and web UI. The interesting part is the decision pipeline: it aggregates predictions across test-time augmentations, uses prediction entropy to decide when to consult a reference embedding bank, and applies kNN voting to refine uncertain predictions. It’s a prototype, and I’d love feedback on the evaluation: • How would you compare classifier-only, always-on kNN, and entropy-gated kNN on the same detected crops? • How would you measure rare-defect recall while separating detection misses from classification errors? Code and examples: If you work on industrial vision or retrieval-based classification, what would you test first?