Helpful OpenCV Tutorials
September 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 moreSeptember 9, 2026 · SentientSight Team
How to run a foundation monocular depth model inside an OpenCV 5 pipeline: ONNX export, DNN preprocessing, converting relative output into metres, temporal stability, honest error metrics, and when one camera is the wrong sensor.
Read moreSeptember 9, 2026 · SentientSight Team
A production 2D pose pipeline on OpenCV 5: top-down vs one-stage models, ONNX export and letterboxing, per-keypoint confidence gating, One-Euro smoothing per track, ground-plane homography for honest zone logic, and event-level precision/recall you can measure.
Read moreSeptember 9, 2026 · SentientSight Team
A robot-cell tutorial for OpenCV 5: frame conventions, a hand-eye capture protocol that is actually solvable, calibrateHandEye and its validation residual, solvePnPRansac to a base-frame grasp pose, and the error budget behind a millimetre spec.
Read more