OpenCV

Call cv2.dnn.readNet("model.onnx") (or readNetFromONNX), build a blob with cv2.dnn.blobFromImage, pass it in with net.setInput(blob), then call net.forward() to get the output array. No training framework is needed at runtime.

detectAndCompute returns a two-element tuple: a list of cv2.KeyPoint objects and a NumPy descriptor array with one row per keypoint. Row i describes keypoints[i]. When nothing is detected, the descriptor value is None, not an empty array.

Use cv2.INTER_AREA. It averages over the source pixel area that maps to each output pixel, so detail is integrated rather than skipped. The default cv2.INTER_LINEAR samples only a few points and produces aliasing and moire on large reductions.

cv2.imread() returns pixels in BGR channel order, while matplotlib's imshow expects RGB. The first and third channels are interpreted swapped, so reds render as blues. Convert first with cv2.cvtColor(img, cv2.COLOR_BGR2RGB).

MORPH_OPEN erodes then dilates: it removes small white specks and thin bridges. MORPH_CLOSE dilates then erodes: it fills small holes and joins nearby blobs. Both roughly preserve blob size, unlike a bare erode or dilate.

OpenCV then ignores the threshold value you passed and computes one from the image histogram by maximizing between-class variance. The chosen value comes back as the first return value of cv2.threshold, which is why the call is written with 0 as the threshold argument.

cv2.dnn.blobFromImage resizes the image to the network's input size, subtracts a per-channel mean, multiplies by scalefactor, optionally swaps the red and blue channels, and returns a 4-D NCHW float array ready for setInput.

Read the model's output layout, keep rows above a confidence threshold, rescale the normalised coordinates by the original frame's width and height, convert them to x, y, w, h, then call cv2.dnn.NMSBoxes to drop overlapping duplicates and draw only the returned indices.

ORB emits binary descriptors — 256 comparison bits packed into 32 uint8 bytes — so distance means counting differing bits, which is what cv2.NORM_HAMMING does. NORM_L2 treats those packed bytes as numeric magnitudes, producing meaningless distances. SIFT's float32 vectors need NORM_L2.

knnMatch with k=2 returns the two nearest descriptors for each query. Keeping a match only when the best distance is below roughly 0.75 times the second-best discards ambiguous matches, where a descriptor fits two candidates almost equally well and is therefore untrustworthy.

cv2.GaussianBlur weights neighbours only by distance, so it averages straight across edges. cv2.bilateralFilter multiplies that spatial weight by a second weight based on intensity difference, so pixels on the far side of an edge contribute almost nothing.

They drive hysteresis. Pixels with gradient magnitude above the upper threshold are accepted as strong edges; below the lower they are discarded; in between they are kept only if they connect to a strong edge. A common ratio is 1:2 to 1:3.

For an 8-bit image, OpenCV stores hue as 0-179, halving the usual 0-360 degrees so it fits in a byte; saturation and value use 0-255. Copying hue values from a colour picker that reports 0-360 selects the wrong colour band.

cv2.calcOpticalFlowPyrLK is sparse: you give it specific points and it returns where each moved, plus a status array. cv2.calcOpticalFlowFarneback is dense: it returns a displacement vector for every pixel. Choose sparse when you track a few features cheaply, dense when you need motion everywhere.

MOG2 models each pixel's recent history as a mixture of Gaussians and returns an 8-bit mask: 0 for background, 255 for foreground. With detectShadows=True (the default) it adds a third label, 127, for pixels it judges to be shadow — so a naive nonzero test counts shadows as objects.

VideoWriter.write() silently ignores any frame whose size or channel count differs from the frameSize and isColor passed to the constructor, and the writer never opens at all if the fourcc codec is unavailable. Check isOpened() and pass (width, height).

Pass matched point pairs as (N,1,2) float32 arrays with method cv2.RANSAC and a reprojection threshold in pixels. The call returns the 3x3 homography plus an inlier mask; the inlier count from that mask, not the raw match count, is the signal that the fit is trustworthy.

Salt-and-pepper noise is isolated extreme pixels. cv2.medianBlur takes the middle value of each neighbourhood, so an outlier is discarded outright. cv2.GaussianBlur averages the outlier in, spreading one bad pixel into a visible grey smudge.

cv2.imread() returns None on any failure — missing path, unreadable permissions, corrupt data, or an unsupported format. It never raises. The error surfaces later, usually as 'NoneType' object has no attribute 'shape', so check for None immediately after the read.

When illumination varies across the frame. cv2.adaptiveThreshold computes a separate threshold for every pixel from the mean of its blockSize x blockSize neighbourhood, minus a constant C, so a shadow or lighting gradient cancels out instead of blacking out half the image.

Two routes. Per contour: cv2.moments gives m00 as the area and m10/m00, m01/m00 as the centroid. Per labelled region: cv2.connectedComponentsWithStats returns stats with CC_STAT_AREA in pixels plus a centroids array - remember label 0 is the background.

Called with no argument, cv2.dnn Net.forward returns a single blob — the first output. A multi-head detector such as a Darknet YOLOv3 net has several output layers, so you must pass net.getUnconnectedOutLayersNames() to forward and receive a list of blobs.

cv2.cornerHarris returns a float32 response map the same size as the input image, which you must threshold yourself. cv2.goodFeaturesToTrack scores corners with the Shi-Tomasi minimum-eigenvalue rule and returns an already-filtered array of corner coordinates.

A derivative is signed: bright-to-dark transitions give negative values. CV_8U cannot hold negatives, so they saturate to 0 and one polarity of every edge vanishes. Compute into CV_16S/CV_32F/CV_64F, then convert with cv2.convertScaleAbs.

The default flag cv2.IMREAD_COLOR always yields a 3-channel 8-bit BGR array: it drops any alpha channel and truncates 16-bit files to 8 bits. cv2.IMREAD_UNCHANGED returns the data as stored, alpha and bit depth included.

cv2.split() allocates a separate copy of every channel, so use it only when you need standalone single-channel arrays. NumPy indexing like img[:, :, 0] returns a view with no copy — cheaper for reading, and writes through to the original image.

Both are single-object trackers seeded once with a box: create with cv2.TrackerCSRT.create() or cv2.TrackerKCF.create(), call init(frame, bbox), then update(frame) each frame for (success, bbox). CSRT is more accurate and handles scale change better; KCF is considerably faster.

Hierarchy has shape (1, N, 4); each row is [Next, Previous, First_Child, Parent] holding indices into the contours list, with -1 meaning none. RETR_TREE is needed when nesting matters - counting holes, or telling an outer boundary from the shapes inside it.

setPreferableBackend and setPreferableTarget express a preference, not a requirement. If the binary was not built with CUDA support — pip's opencv-python wheels are not — OpenCV logs a warning and silently falls back to the default CPU implementation, so timings stay flat.