Glossary

Every term the docs use, in plain language, with a link to where it's put to work. Skim it once and the rest of the site reads faster.

Images & pixels

Pixel — one dot of an image; a number (greyscale) or a few numbers (colour). See Image basics.
Channel — one of the numbers a pixel carries. 1 = greyscale, 3 = colour, 4 = colour + transparency.
BGR — OpenCV's colour-channel order: Blue, Green, Red — not RGB. The classic beginner gotcha; Image basics explains it.
Greyscale — a one-channel image; each pixel is a single 0–255 intensity. Many operations (edges, thresholds) want greyscale input.
HSV — Hue/Saturation/Value colour space. Thresholding by hue is far more robust than by RGB for "find the green things" — see Colour masking.
Alpha channel — a 4th channel storing per-pixel transparency (0 = clear, 255 = opaque).
Data type / depth — how each channel value is stored: CV_8U (0–255, the everyday one), CV_16S (signed, for derivatives), CV_32F (float, for maths). See Image basics.
ROI (Region Of Interest) — a rectangular sub-area of an image you want to work on.

Mat — OpenCV's native image/matrix: the off-heap grid of pixels. You rarely touch it directly; Image wraps it.
Image — scalacv's high-level, owned image you transform by chaining. The tier to reach for first.
Managed[A] — a handle that frees its native object exactly once and throws (not segfaults) if used after release. The heart of the memory model.
Scalar — up to four channel values in one value — a colour, in BGR order. Named constants: Scalar.Red, Green, Blue, Black, White.
Point / Size / Rect — 2-D geometry: a coordinate (x, y), an extent width × height, and a rectangle. See Geometry.
CvError — scalacv's typed error hierarchy for expected, data-dependent failures (DecodeFailed, LoadFailed, NativeCall, …). See The error model.

Native memory — memory allocated by C++ (the pixel buffers), off the JVM heap. The garbage collector can't see the pressure, which is why release must be explicit — see Mat lifecycle.
Move semantics — an Image transform consumes the image it's called on and returns a new one, so a chain holds one live buffer at a time. Reusing a consumed image throws.
Owned / borrowed / copied-out — the three ownership dispositions: you close owned handles; you must not close borrowed ones; copied-out results (a Contour, a Rect) are plain data you keep forever. See Architecture.
RSS (Resident Set Size) — the process's real physical memory. The only reliable signal for a native leak, since Pointer.totalBytes() can't see OpenCV's buffers — see Performance.

Kernel — a small grid of weights slid over an image to blur, sharpen, or detect edges. Blur/morphology sizes are kernels.
Threshold — turn a greyscale image into black-and-white by a cutoff; adaptive threshold computes the cutoff per-neighbourhood for uneven lighting. See Image processing.
Morphology — shape operations on binary images: erosion shrinks bright regions, dilation grows them; opening/closing combine them.
Canny — the classic edge detector; outputs a one-channel edge map. See Image processing.
Contour — the outline of a connected blob, as a list of points; the output of findContours. See Contours.
Histogram — a count of how many pixels fall in each intensity bucket; equalising it stretches contrast.
Inpainting — filling a masked-out region from its surroundings, to erase a scratch or object. See the Cookbook.
Hough transform — detects straight lines in an edge image. scalacv wraps the line variants only: houghLines returns infinite lines as PolarLines, houghLinesP returns finite Segments with real endpoints. OpenCV's circle variant, HoughCircles, is not wrapped — to find round shapes, take contours() and measure circularity (4 * math.Pi * c.area / (c.perimeter * c.perimeter), which is 1.0 for a perfect circle and falls away as the outline gets less round), or call Imgproc.HoughCircles yourself through the raw-Mat escape hatch. See Hough.

Feature / keypoint — a distinctive, repeatably-findable spot in an image (a corner, a blob). Descriptors encode the look around a keypoint so it can be matched across images. See Object detection.
Cascade (Haar/LBP) — a fast, classic object detector defined by an XML file (bundled, nothing to download). Good for faces. See Object detection.
DNN — Deep Neural Network. scalacv runs pre-trained networks via OpenCV's dnn module. See DNN inference.
Blob (image) — a connected region of foreground pixels in a binary image: the white patch left behind after a threshold or a colour mask, and the thing contours() outlines. This is the sense the tutorials use when they talk about counting blobs. See Contours.
Blob (DNN) — an unrelated use of the same word, from a different corner of OpenCV: the pre-processed 4-D tensor (batch × channels × height × width) fed into a network. It only turns up around Dnn.blobFromImage and the DNN page.
ONNX — an open model format; the usual way to bring a trained network into scalacv.
Optical flow — how pixels/features move between consecutive frames; the basis of motion and odometry.
ArUco marker — a printed square barcode used as a known reference for augmented reality and pose.

Intrinsics — a camera's internal parameters (focal length, principal point, lens distortion). Turns pixels into rays. See Calibration.
Calibration — measuring a camera's intrinsics from photos of a known target (a chessboard).
Pose — an object's or camera's 3-D position + orientation (rotation + translation). See Pose estimation.
Homography — the 3×3 transform mapping one plane to another (e.g. a marker's flat face to the image).
Kalman filter — a predictor/smoother that tracks a moving target through noise and gaps. See Tracking.
SLAM — Simultaneous Localization And Mapping: figuring out where the camera is and building a map as it moves. scalacv provides the front end (odometry, loop closure); see Navigation.
Loop closure — recognising a place the camera has already visited, so a map can correct its drift.

FOURCC — a four-character code naming a video codec (MJPG, mp4v). See Video.
PSNR / max-abs-diff — tolerance metrics for comparing images without demanding bit-for-bit equality (which SIMD/platform differences break). See Testing.
mdoc — the tool that compiles every code snippet in these docs against the real library, so no example can drift out of date.