Helpful OpenCV Tutorials
September 9, 2026 · SentientSight Team
Every rescued vision project starts with “it worked last month”. Pinning the whole stack, building a representative golden-frame corpus, tolerance-based assertions instead of pixel equality, per-slice model-promotion gates, containerised CI, and production drift monitoring that closes the loop back into the test suite.
Read moreSeptember 9, 2026 · SentientSight Team
A measurement-first guide to OpenCV 5 performance: building a per-stage frame-time budget, reading cv::TickMeter and py-spy honestly, when UMat/OpenCL and CUDA actually pay off, hiding host-device copies, G-API graphs, and the threading mistakes that silently halve throughput.
Read moreSeptember 9, 2026 · SentientSight Team
A production scene-text and industrial code-reading pipeline in OpenCV 5: DB text detection, perspective-correct crops, CRNN/CTC recognition, lighting and preprocessing fixes, check digits, temporal voting, and read-rate vs error-rate acceptance testing.
Read moreSeptember 9, 2026 · SentientSight Team
How to prototype a detector from a text prompt instead of a labelled dataset: exporting YOLO-World to ONNX, running it in OpenCV 5's DNN engine, promoting boxes to masks with SAM 2, and the honest limits before you ship it to production.
Read moreSeptember 9, 2026 · SentientSight Team
Why a 94% mAP detector still produces a 30% wrong count: reliable GStreamer RTSP ingest, a ByteTrack-style association loop in plain NumPy, foot-point line and zone counting, ReID for occlusions, and HOTA/IDF1 evaluation.
Read more