Helpful OpenCV Tutorials
September 18, 2026 · SentientSight Team
Python 3.14's free-threaded build is officially supported, and it changes how a multi-camera OpenCV pipeline should be architected. How to check your wheels, where cv2 is and is not thread-safe, a worker-pool pattern that scales on free-threaded CPython, and how to measure whether it actually helped.
Read moreSeptember 17, 2026 · SentientSight Team
Linking the same person or pallet across cameras: why geometry beats embeddings, validating a ground-plane homography in metres, clock sync, ONNX re-ID crops as a gated tie-breaker, and the handover precision/recall numbers to quote.
Read moreSeptember 16, 2026 · SentientSight Team
Why cv2.Stitcher fails on inspection data, how to hand-build the pipeline with the right motion model, USAC_MAGSAC and SuperPoint/LightGlue via ONNX, and how to turn a mosaic into millimetres you can put in an acceptance test.
Read moreSeptember 15, 2026 · SentientSight Team
The accuracy problem is usually an acquisition problem. Manual exposure and white balance, flat-field correction, Mertens exposure fusion, low-light denoising that does not destroy the signal, and a frame-quality gate that rejects bad input instead of quietly degrading your metrics.
Read moreSeptember 14, 2026 · SentientSight Team
Where a small vision-language model (Qwen2.5-VL, Florence-2, Moondream) genuinely helps a production vision system, where it must never be trusted, and how to wire one behind an OpenCV 5 pipeline with constrained JSON output, a confidence gate, cost caps and a regression harness.
Read more