Glossary

The specific arrangement of layers and connections that defines one named neural network design.

An explicit scene representation using many overlapping 3D Gaussian blobs, rendered by fast splatting instead of ray-marching.

The fraction of predictions a classifier gets right -- a metric that misleads badly when classes are imbalanced.

How well a backbone's features perform on your actual target task, not just on whatever benchmark a paper originally reported.

Giving every pixel its own black-or-white cutoff, based on the small patch of pixels around it, instead of one cutoff for the whole image.

A linear map plus a translation -- preserves parallel lines but not necessarily angles or lengths.

The false, coarser pattern a sampling grid records when a repeating scene detail is finer than the grid can tell apart from a slower one.

A predefined reference box shape and size, tiled across the image, that a one-stage detector predicts offsets from.

The adjustable opening that controls how much light reaches a camera's sensor.

A search structure that skips comparing most gallery embeddings, trading a small, measured accuracy loss for speed.

Combining every region's recognized text into one final structured result, in the order the document's own structure calls for.

The shared feature-extraction network that one or more task-specific heads are attached to.

The ordered set of questions -- latency, then labeled-data scarcity, then accuracy -- used to choose a backbone for a real project.

The algorithm that computes how much to nudge every weight, working backward from the loss.

A fixed-length feature vector built by clustering local descriptors into a vocabulary and histogramming which cluster each descriptor falls into.

The physical distance between the two camera centers in a stereo pair.

An extra leading axis on an image tensor counting how many images are stacked together.

The color-filter mosaic over a sensor's pixels, and the interpolation that reconstructs a full-color image from it.

Forcing every pixel in an image down to pure black or pure white, usually to separate ink from paper.

A compact bit-string keypoint descriptor built from pairwise pixel-intensity comparisons, much cheaper to match than a gradient histogram.

The smallest axis-aligned rectangle that contains a shape.

Adjusting every camera pose and every 3D point together so predicted and observed pixels agree as closely as possible.

Estimating a camera's intrinsic and extrinsic parameters from known point correspondences.

The rotation and translation describing where a camera is and which way it's pointed in the world.

The parameters describing a camera's own internal geometry -- focal length, principal point, and skew -- independent of where it's placed.

Whether a tensor's color-channel axis comes right after the batch axis or comes last.

Triangulating each candidate camera pose and keeping only the one that puts a point in front of both cameras.

A vision system built from hand-engineered features feeding a general-purpose classifier, not a neural network.

A trained model that assigns one of a fixed set of labels to an input.

A shared embedding space for images and text, trained by contrastive learning on paired image-caption data.

Getting a machine to extract meaning -- what, where, how far -- from an image or video.

A detector's own number, between 0 and 1, for how sure it is that a predicted box actually contains the object.

A table sorting every prediction into true positive, false positive, false negative, or true negative, once compared against ground truth.

The traced outer boundary of a connected region in an image.

Replacing a dense boundary trace with a small set of vertices that reproduce its shape within a tolerance.

A training objective that pulls matching pairs of embeddings together and pushes mismatched pairs apart.

The smallest convex polygon that contains a shape's entire boundary.

An algorithm that flags image locations whose local gradient structure looks like a corner.

Sliding a small weight grid over an image and summing at every position; convolution flips the grid first.

A matched pair of points -- one from each of two images -- believed to show the same physical point.

A similarity score between two vectors based on the angle between them, ignoring their magnitude.

Applying label-preserving transformations to training images to synthetically expand a dataset and reduce overfitting.

A component that produces output text one token at a time, conditioning each new word on the input and on the words already generated.

A vision system that learns its own feature representation end-to-end from data, typically via a neural network.

Getting a trained model running on the actual hardware it will be used on, not just in a training notebook.

Pairing each keypoint descriptor in one image with its closest-distance match in another.

Rotating a photographed or scanned page back to level, undoing the small tilt a camera or scanner almost always adds.

A mask-overlap metric that weights shared area twice against the sum of both masks' areas.

A generative model trained to reverse a gradual noising process, generating an image by denoising from pure noise.

The horizontal pixel-position difference between a matched point's location in a rectified stereo pair.

Grouping a document's detected and recognized text into structured regions -- paragraphs, headers, tables.

A second (or third) box a detector reports for the same real object, rather than a distinct one.

One eigenvector of an image dataset's covariance matrix, reshaped back into picture form.

A number and matching direction that a matrix only stretches, never rotates.

The trained network that turns a raw image into its embedding vector.

The geometric relationship between two camera views that constrains a matching point to lie on a single line.

The single point every epipolar line in an image passes through -- the other camera's own center, as seen in this image.

Ordinary straight-line distance between two points -- the square root of the sum of squared coordinate differences.

The three settings -- aperture, shutter speed, and ISO -- that jointly determine an image's exposure.

The harmonic mean of precision and recall, a single number summarizing both at once.

A fixed-length numeric summary of an image, built from shape/texture/color/keypoint statistics, used as a classifier's input.

The angular extent of the scene a camera actually captures, set jointly by focal length and sensor size.

Continuing to train some or all of a pretrained model's own weights on a new task, usually at a low learning rate.

Estimating a derivative by comparing a pixel to its neighbors instead of using calculus.

The lens-to-sensor distance that sets how much a scene is magnified and how wide a view the camera captures.

The pixels belonging to the object of interest (foreground) versus everything else (background).

The way a camera's perspective visually compresses parts of a scene farther from the image center or farther from the camera.

Feeding an input through every layer of a network in order to produce its final output.

The stored collection of candidate images (in practice, their embeddings) that a query gets compared against.

Two networks -- a generator and a discriminator -- trained against each other, the generator learning to fool the discriminator.

A count of how many pixels in a patch have their gradient pointing in each of several directions.

Representing a 2D or 3D point with one extra coordinate so translation becomes a matrix multiply like every other transform.

The 3x3 projective transform matrix mapping points on one plane to their image on another.

One of seven shape descriptors built from image moments that stay the same under translation, scale, and rotation.

A learned fixed-length vector representation of an image's content, placed in a latent space where distance tracks similarity.

Producing a new image from a description, a sample, or another image, rather than interpreting one that already exists.

An arrow at each pixel pointing toward the direction brightness increases fastest, sized by how fast it increases.

A weighted sum of a shape's pixel coordinates that summarizes its size, position, or orientation numerically.

Transforming an image into another image, as distinct from computer vision's goal of extracting meaning from one.

The end-to-end task of finding the gallery images most similar to a query image.

Extracting meaning from an existing image -- what it shows, where, how far away.

A large, standard benchmark dataset of labeled photos widely used to pretrain and evaluate vision models.

The real wall-clock time a model takes to run one forward pass on the hardware it will actually run on.

Filling in a masked or missing region of an image (inpainting), or extending an image past its original edges (outpainting).

Labeling every pixel by class and separating each individual object instance from every other.

The overlap between a predicted region and the ground truth, divided by their combined area.

A descriptor's ability to come out the same number even after the photo is scaled, rotated, or relit.

One random-sample-and-fit attempt in RANSAC's loop -- pick a minimal sample, fit, count agreement, then try again.

A recursive predict-then-correct estimator that tracks an object's state frame-by-frame under uncertainty.

An image location a detector flags as distinctive enough to find again reliably.

A real lens's deviation from the ideal pinhole model, mainly radial (barrel/pincushion) and tangential.

A transform whose output is only a weighted mix of the input's own numbers -- no shifting, no extra constant added.

Recognizing that a camera has returned to an already-mapped place, and using that fact to correct drift across the whole path.

A grid showing how strongly an image repeats at each possible speed and direction.

A two-stage detection network extended with a branch that predicts a segmentation mask for each detected object.

A detector's average precision, averaged again across every object class, into one summary number.

Reducing a trained model's numeric precision (e.g. 32-bit floats to 8-bit integers) to shrink its memory and latency footprint.

Estimating per-pixel depth from a single image, with no second camera view to triangulate from.

Maintaining a consistent identity for each of several tracked objects across a video, including through brief occlusion.

Retrieving the closest embeddings to a query embedding from a gallery, by a similarity or distance metric.

An implicit scene representation -- a trained network that outputs color and density for any 3D point and viewing direction, enabling novel-view rendering.

A unit that multiplies its inputs by weights, adds a bias, sums, and applies a nonlinearity.

Collapsing a detector's many overlapping candidate boxes for the same object down to one, keeping the highest-confidence box.

Converting a trained model into the framework-independent ONNX format for portable inference deployment.

The per-pixel apparent-motion field estimated between two consecutive video frames.

A model that memorized its training data (overfitting), versus one too simple to fit even that data (underfitting).

A single wide photo built by aligning and blending several overlapping photos taken from one fixed viewpoint.

Two lines that point in exactly the same direction and never meet, however far they're extended.

A Vision Transformer's first step -- splitting an image into fixed-size patches and linearly projecting each into a token vector.

The five-stage spine -- image, preprocess, features, model, prediction -- that any vision system can be placed on.

A traced contour's own length -- the sum of the straight-line distance between each consecutive pair of boundary points.

Dividing a homogeneous point's first coordinates by its last one to recover its true image-plane position.

The idealized camera model of light passing through a single point onto an image plane.

A set of 3D points, typically with no connectivity between them, representing a scene's geometry.

Recovering a camera's own unknown rotation and translation from point correspondences alone.

A specific, labeled body-joint location (e.g. left elbow, right knee) detected in an image or frame.

The two pipeline stages a deployed system adds right after Model -- cleaning up a raw prediction, then displaying it for a person to see.

Of everything a model flagged as positive, the fraction that was actually correct.

Where the camera's optical axis pierces the image sensor -- the sensor's own optical center, not necessarily its pixel-grid center.

A number between 0 and 1 measuring how likely an event is, with 1 meaning certain.

A more general transform than affine that can converge parallel lines to a vanishing point.

Segmenting an object from a point, box, or text prompt supplied at inference time, rather than a fixed trained label set.

Lens distortion that pushes a point along the line from the image center through it, by an amount that grows with distance from the center -- outward (barrel) or inward (pincushion).

A robust model-fitting procedure that repeatedly samples minimal point sets and keeps whichever fit gets the most support.

Rejecting a descriptor match unless its best candidate is meaningfully closer than its second-best.

Of everything that was actually positive, the fraction the model actually found.

The region of the original input image that a given unit's activation is actually influenced by.

Warping a stereo image pair so corresponding epipolar lines line up as the same horizontal row in both images.

The three properties a keypoint must have at once -- findable again, tellable apart from its surroundings, and pinned to an exact spot.

The distance between a point's observed image location and where a calibrated camera model predicts it should land.

A separate, unweighted path that carries a block's input forward and adds it back to the block's output, so the block only has to learn a correction.

A transform that turns every point the same angle around a fixed center, changing no distances.

A mechanism that lets every token in a sequence compare itself against every other token and blend their information by a learned weight.

Learning useful features from unlabeled images alone, with no labels at all during pretraining.

Labeling every pixel in an image by class, with no notion of separate object instances.

A gradient-histogram feature vector built around a keypoint to make matching invariant to scale, rotation, and illumination.

Tracking a moving camera's own pose while building a map of the scene, from a live stream, at the same time.

Two cameras viewing the same scene together, so their images can be combined to recover depth.

Recovering both a scene's 3D structure and every camera's pose from many overlapping images with no camera positions known in advance.

Lens distortion caused by the lens not sitting perfectly parallel to the sensor, corrected with its own two coefficients.

Locating the regions of an image that contain text, without reading what they say.

Converting the pixels inside a detected text region into an actual character string.

The three disjoint pieces a labeled dataset is split into before training, each with its own job.

Reusing a model's weights, pretrained on one task, as the starting point for a new task by retraining only part of it.

Sliding every point of a shape by the same fixed amount in the same direction.

Recovering a 3D point's location by intersecting the two camera rays that produced its matched image points.

The single point in an image where lines that are parallel in real life appear to converge because of perspective.

A generative model pairing an encoder that maps inputs to a latent distribution with a decoder that reconstructs from samples of it.

One very large model, pretrained once on huge amounts of data, that can be pointed at many different vision tasks through a short prompt instead of being retrained from scratch.

A vision model that treats an image as a sequence of patch tokens and processes it with self-attention instead of convolution.

A model that jointly processes images and text -- captioning, visual question answering, or retrieval.

Estimating a camera's own motion between consecutive video frames using only the images -- no GPS, no external tracking.

Classifying an image with no task-specific training, by comparing its embedding to a set of text-prompt embeddings.

Segmenting every instance of a concept the model was never specifically trained to segment.