Helpful OpenCV Tutorials
September 28, 2026 · SentientSight Team
Why stalled computer vision projects are almost always data problems: pHash and embedding dedup in OpenCV 5, annotation guides with a measured agreement ceiling, model pre-labelling, an uncertainty-plus-diversity active-learning loop, and a frozen golden test set that tells you when to stop paying for labels.
Read moreSeptember 21, 2026 · SentientSight Team
How to wire an OpenCV 5 pipeline into a ROS 2 Jazzy/Kilted robot stack without paying for it in copies: cv_bridge without the ABI surprises, intra-process composition, image_transport compression, deadline QoS, and a latency budget you can quote to a customer.
Read moreSeptember 21, 2026 · SentientSight Team
AI PC and edge NPUs are now the cheapest frames-per-watt you can buy, but they only accept quantised models. A practical OpenCV 5 workflow: ONNX export hygiene, post-training INT8 with a real calibration set, per-layer fallback when accuracy drops, NPU vs GPU vs CPU benchmarking, and the accept/reject report that keeps a client sign-off honest.
Read moreSeptember 21, 2026 · SentientSight Team
When the PyPI wheels stop being enough: an explicit CMake configuration, a multi-stage CUDA Dockerfile, a getBuildInformation() self-check that fails the build, realistic image sizes, NumPy/ABI traps, and the provenance bundle you hand to a client.
Read moreSeptember 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 more