Coming from Python

Examples use named imports after one await initOpenCV() call. See the initialization guide.

The algorithms are compiled from OpenCV 5 C++. The JavaScript interface follows native argument order, destination matrices and explicit resource lifetime. Python names are widely available, but NumPy behavior and Python calling conventions are not emulated.

Python pattern	TypeScript pattern
gray = cv2.cvtColor(image, code)	using gray = new Mat(); cvtColor(image, gray, code)
np.array(...)	matFromArray(rows, cols, type, typedArray)
Array slicing for an image region	Native ROI or an explicit JavaScript copy, with attention to row stride
Returned tuples	Named result fields, such as getTextSize(...).value and .baseLine
Garbage-collected NumPy arrays	using or explicit delete() for native handles
Keyword arguments	Positional arguments from the exact TypeScript overload
cv2.imread(hostPath)	Read bytes yourself and decodeImage, or put a file in FS

Factories and namespaces

Both familiar aliases and native class factories are typed. OpenCV 5 also moves some feature algorithms into xfeatures2d; inspect the actual reference instead of assuming the OpenCV 4 layout.

import { KeyPointVector, Mat, SIFT_create } from '@banou/opencv-wasm'
using detector = SIFT_create()
if (!detector) throw new Error('SIFT was not created')
using points = new KeyPointVector()
using descriptors = new Mat()
using noMask = new Mat()
detector.detectAndCompute(image, noMask, points, descriptors)

Overload families

Some native overload groups have separate names, such as findHomography1, train1 or from1. The suffix identifies a binding family, not a quality level. Read its signatures to see the accepted types and returned fields.

Preserve numeric intent

JavaScript arithmetic uses double precision. Assigning to a Float32Array rounds to float32; Math.fround can reproduce an intermediate float32 operation when a migration needs it. NumPy integer casts, OpenCV saturated conversions and JavaScript typed-array casts do not have identical behavior.

For a faithful port, compare masks, shapes, coverage and numeric outputs with native fixtures. Small floating-point differences can cross a discrete threshold and alter a connected region or crop boundary. A name-coverage percentage cannot establish this behavior.

See compatibility for the measured Python inventory.