OpenCV

OpenCV (Open Source Computer Vision Library) is a widely used open-source library of classical computer-vision algorithms — image filtering, edge detection, feature matching, camera calibration — written in C++ with bindings for Python and other languages. It predates the deep learning era and is distinct from a CNN: OpenCV's classical algorithms are hand-engineered, not learned, though it's commonly used alongside deep learning models for preprocessing images (resizing, color conversion) or running inference.

How it works

The library is a C++ core with thin language bindings; in Python everything hangs off cv2 and images are plain NumPy arrays. Two details trip up nearly everyone: cv2.imread returns channels in BGR order rather than RGB, and arrays are indexed (height, width, channels) while most function arguments take sizes as (width, height). The algorithms themselves are deterministic and hand-designed — cv2.Canny for edges, cv2.warpAffine for geometric transforms, ORB or SIFT for keypoint matching, cv2.VideoCapture for stream I/O. The dnn module additionally loads trained models from ONNX and similar formats, which makes OpenCV a dependency-light path to inference when a full PyTorch install is unwarranted.

When it breaks

• Color-space bugs are silent: feeding BGR arrays to a model trained on RGB costs accuracy without raising anything.
• Failures surface late. cv2.imread returns None for a missing or unreadable file instead of raising, so the real error appears downstream as a confusing NoneType crash.
• Classical feature detection and matching degrade sharply under lighting changes, motion blur, and low-texture surfaces — exactly the cases learned features handle better.
• The dnn module trails mainstream frameworks in operator coverage and GPU support; treat it as a deployment convenience, not a training stack.

See also: CNN

Learn more: Computer Vision · OpenCV official docs