Computer vision glossary

161 terms defined in plain English — the vocabulary that turns up in computer vision job descriptions, interviews and technical discussions. Where a term has a fuller page on the site, each entry links to it.

Fundamentals

Colour spaceRGB · HSV · YUV · Lab: A coordinate system for representing colour. RGB is the capture and display default, HSV separates hue from brightness for thresholding, YUV separates luminance from chrominance for compression, and Lab approximates perceptual uniformity.; See also: Bayer pattern & demosaicing, ISP
Computer visionCV: The field concerned with extracting usable information from images and video — what is present, where it is, how it is moving, and what it means. It spans classical signal processing and modern deep learning, and in industry almost always includes the data and deployment engineering around the model.; See also: Machine vision, Image processing
Convolution: Sliding a small kernel across an image and computing a weighted sum at each position. It is the basic operation of both classical filtering and convolutional neural networks — the difference is only whether the kernel is designed or learned.; Filter design jobs →See also: Kernel, CNN, Receptive field
Histogram equalisationCLAHE: Redistributing pixel intensities to improve contrast. CLAHE — contrast-limited adaptive histogram equalisation — does this in local tiles with a clipping limit, and is widely used as preprocessing in medical and low-light imaging.; See also: Image processing
Image processing: Operations that transform an image into another image — filtering, denoising, sharpening, colour correction. Distinguished from computer vision proper, which aims to produce an interpretation rather than a picture, though the two are used together constantly.; Image enhancement jobs →See also: Convolution, ISP
KernelFilter · Structuring element: The small matrix of weights applied during a convolution. Classical kernels such as Gaussian, Sobel and Laplacian have known effects; in a neural network the kernel weights are learned from data.; See also: Convolution
Machine vision: Computer vision applied to industrial automation — inspection, measurement and guidance on production lines. The term implies controlled lighting, fixed cameras and hard throughput requirements, which is why classical deterministic methods remain common.; Industrial inspection jobs →See also: Computer vision, Morphological operations
Morphological operationsErosion · Dilation · Opening · Closing: Set-theoretic operations on image shape — eroding, dilating, opening and closing regions using a structuring element. Fast, deterministic and still the standard way to clean up a binary mask or separate touching objects.; Morphological processing jobs →See also: Segmentation, Watershed transform
Receptive field: The region of the input image that influences a particular output value. Deep convolutional networks build large receptive fields by stacking layers; transformers have a global receptive field from the first layer, which is much of why they behave differently.; See also: Convolution, Vision transformer
Watershed transform: A segmentation technique that treats intensity as elevation and floods the image from markers, with boundaries forming where floods meet. The classic method for separating touching objects such as cells or packed components.; See also: Morphological operations, Segmentation

3D & Geometry

3D Gaussian splatting3DGS: Representing a scene as millions of anisotropic 3D Gaussians rasterised directly to the screen. Achieves NeRF-quality novel views while training far faster and rendering in real time, which is why it displaced NeRF in most products.; See also: NeRF, Scene reconstruction
6-DoF pose: The full position and orientation of a rigid object — three translation and three rotation degrees of freedom. What robotic grasping and AR object placement need, as distinct from 2D or 3D keypoint pose.; Pose estimation jobs →See also: PnP, Keypoint
Bird's-eye viewBEV: A top-down representation into which features from multiple cameras are projected. Fusion across sensors and time is natural in this space, and planning operates there anyway, which is why BEV became the standard intermediate representation in modern autonomy stacks.; 3D object detection jobs →See also: Occupancy grid, Sensor fusion, Homography
Bundle adjustmentBA: Jointly refining camera poses and 3D point positions to minimise reprojection error across all views. The computational heart of structure from motion and the back end of most SLAM systems, usually solved with Ceres or g2o.; Ceres Solver jobs →See also: Reprojection error, Structure from motion, Pose graph
Calibration: The process of determining intrinsics and extrinsics, typically using a known target such as a checkerboard or ChArUco board. At fleet scale it becomes its own engineering discipline covering factory procedure, drift monitoring and online recalibration.; Hiring calibration engineers →See also: Camera intrinsics, Camera extrinsics, Hand-eye calibration
Camera extrinsics: The rigid transform — rotation and translation — relating a camera to something else: another camera, a LiDAR, a robot base or the world. Errors here are the most common cause of multi-sensor systems that appear to work but produce wrong geometry.; See also: Camera intrinsics, Calibration, Coordinate frame
Camera intrinsics: The parameters describing the camera itself — focal length, principal point and lens distortion — which determine how 3D points project onto the sensor. Without them no measurement from an image is metrically meaningful.; Camera calibration jobs →See also: Camera extrinsics, Calibration, Lens distortion
Coordinate frametf · Transform tree: A reference system that positions are expressed in. Real systems maintain a tree of frames — sensor, robot base, odometry, map, world — and a large share of practical perception bugs are frame or timing errors rather than algorithmic ones.; ROS 2 jobs →See also: Camera extrinsics, Sensor fusion
Depth estimation: Predicting per-pixel distance from stereo, monocular or sequential images. Monocular depth is fundamentally scale-ambiguous — a small near object and a large far one project identically — so metric scale must come from elsewhere.; Depth estimation jobs →See also: Stereo vision, Disparity, Monocular
Disparity: The horizontal pixel offset of a scene point between two rectified stereo views. Inversely proportional to depth — large disparity means close, small means far — which is why distant objects are hard to range accurately.; See also: Stereo vision, Triangulation
Epipolar geometry: The geometric relationship between two views of the same scene. A point in one image constrains its match in the other to a line, which reduces stereo matching from a 2D search to a 1D one. Foundational to essentially all multi-view work.; See also: Fundamental & essential matrix, Stereo vision, Rectification
Features & descriptorsSIFT · ORB · SuperPoint · Keypoints: Distinctive image points plus a description of their local appearance, so the same physical point can be recognised in another view. SIFT and ORB are the classical workhorses; SuperPoint and LightGlue are the learned successors.; Feature extraction jobs →See also: RANSAC, SLAM, Structure from motion
Fundamental & essential matrix: Matrices encoding the epipolar constraint between two views. The fundamental matrix works in pixel coordinates; the essential matrix works in calibrated coordinates and can be decomposed into relative rotation and translation.; See also: Epipolar geometry, Homography, RANSAC
Hand-eye calibration: Solving for the rigid transform between a robot end-effector and a camera mounted on it, classically formulated as AX = XB. A standard requirement in robotic manipulation and a frequent interview question.; See also: Calibration, Camera extrinsics
Homography: A projective transform mapping one plane to another, with eight degrees of freedom. Used for panorama stitching, perspective correction, planar tracking and bird's-eye-view projection of road surfaces.; See also: Fundamental & essential matrix, RANSAC
ICPIterative closest point: Aligning two point clouds by repeatedly matching nearest neighbours and solving for the transform minimising their distance. Underpins registration and LiDAR odometry, and converges to a local minimum so initialisation matters.; See also: Point cloud, Registration
Kalman filterEKF · UKF: A recursive estimator that fuses predictions with measurements under Gaussian assumptions. The extended and unscented variants handle non-linear systems. Still the backbone of tracking and fusion because its uncertainty is calibrated and auditable.; See also: Particle filter, Sensor fusion
Lens distortion: Deviation from the ideal pinhole projection, mostly radial (barrel or pincushion) with a smaller tangential component. Corrected via a distortion model during undistortion or rectification; wide-angle and fisheye lenses need different models entirely.; See also: Camera intrinsics, Rectification
Loop closure: Recognising a previously visited place and using that constraint to correct accumulated drift across the whole trajectory. It is what separates SLAM from odometry, and place recognition under changed lighting is its hard part.; See also: SLAM, Visual odometry, Pose graph
Mesh: A surface representation of connected vertices, edges and faces. The standard output for rendering and simulation, usually extracted from a volumetric reconstruction by marching cubes.; See also: TSDF, Point cloud, Scene reconstruction
Monocular: Using a single camera. Cheaper and simpler than stereo or LiDAR but geometrically ambiguous in scale, so monocular systems rely on learned priors, a second sensor, or known geometry such as camera height above the ground.; See also: Depth estimation, Visual-inertial odometry
NeRFNeural radiance field: Representing a scene as a continuous function from position and view direction to colour and density, rendered by volumetric integration to synthesise novel views. Largely superseded in production by Gaussian splatting.; NeRF jobs →See also: 3D Gaussian splatting, Structure from motion, Scene reconstruction
Occupancy grid: A grid marking cells as free, occupied or unknown. Its appeal in autonomy is that it handles the long tail — a detector trained on cars and pedestrians misses debris, whereas an occupancy grid marks anything solid as not drivable.; Occupancy prediction jobs →See also: Voxel, Bird's-eye view
Particle filterSequential Monte Carlo: Estimating state with a weighted set of hypotheses rather than a single Gaussian. Handles multi-modal beliefs a Kalman filter cannot — a robot that might be in one of three identical corridors — at higher computational cost.; Particle filtering jobs →See also: Kalman filter, SLAM
Photogrammetry: Making metric measurements of the physical world from photographs. Overlaps heavily with structure from motion, but the term implies survey-grade accuracy requirements and is standard in construction, mapping and heritage work.; See also: Structure from motion, Point cloud
PnPPerspective-n-Point: Recovering camera pose from a set of known 3D points and their 2D projections. The core of marker-based tracking, relocalisation and 6-DoF object pose estimation, almost always wrapped in RANSAC.; See also: 6-DoF pose, RANSAC, Triangulation
Point cloud: An unordered set of 3D points, typically from LiDAR, depth cameras or photogrammetry. Its irregular, permutation-invariant structure is why point cloud networks look nothing like image CNNs.; Point cloud processing jobs →See also: Voxel, ICP, LiDAR, Mesh
Pose graph: A graph where nodes are camera or robot poses and edges are relative transform measurements with uncertainty. Optimising it distributes accumulated error consistently, which is the mechanism by which a loop closure fixes a trajectory.; See also: Bundle adjustment, Loop closure, SLAM
RANSACRandom sample consensus: A robust fitting method that repeatedly samples minimal subsets, fits a model, and keeps the one with the most inliers. Essential because feature matching always produces outliers, and a least-squares fit over contaminated data is worthless.; See also: Features & descriptors, Homography, Bundle adjustment
Rectification: Warping a stereo pair so corresponding points lie on the same image row, reducing matching to a 1D search. Depends on accurate calibration, which is why stereo systems need thermal-drift handling and often online recalibration.; See also: Stereo vision, Calibration, Epipolar geometry
Registration: Aligning two datasets into a common coordinate system — two point clouds, two scans, or two medical images taken at different times or with different modalities. Central to both 3D reconstruction and medical imaging.; SimpleITK jobs →See also: ICP, Point cloud
Reprojection error: The pixel distance between where a 3D point actually appears in an image and where the current pose and calibration estimate predict it should. The quantity nearly every geometric optimisation minimises.; See also: Bundle adjustment, Calibration
Scene reconstruction: Building a 3D model of an environment from images or depth data. Spans offline high-fidelity capture and online reconstruction running live on a robot or headset, with the right representation depending on what consumes it.; Scene reconstruction jobs →See also: TSDF, Mesh, NeRF, Structure from motion
Sensor fusion: Combining camera, LiDAR, radar, IMU and GNSS into one coherent estimate. Early fusion preserves the most information but demands tight calibration and synchronisation; late fusion is more robust and modular, which safety-critical systems often prefer.; Sensor fusion jobs →See also: Kalman filter, Coordinate frame, Time synchronisation
SLAMSimultaneous localisation and mapping: Building a map of an unknown environment while tracking your own position within it. The capability that lets a robot, drone or headset operate somewhere it has never been without external positioning infrastructure.; SLAM jobs →See also: Visual-inertial odometry, Loop closure, Visual odometry, Bundle adjustment
Stereo vision: Recovering depth by matching pixels between two calibrated cameras and converting the resulting disparity into distance. Passive, cheap and works in sunlight, but struggles on textureless surfaces and loses accuracy at range.; Stereo vision jobs →See also: Disparity, Rectification, Depth estimation
Structure from motionSfM: Reconstructing 3D geometry and camera poses from an unordered image collection, offline and at high accuracy. Where SLAM prioritises real-time operation, SfM prioritises reconstruction quality; COLMAP is the reference implementation.; Structure from motion jobs →See also: Bundle adjustment, Photogrammetry, NeRF
Triangulation: Recovering a 3D point from its projections in two or more views with known camera poses. The accuracy degrades as the baseline shrinks relative to the distance, which is why stereo depth error grows with the square of range.; See also: Stereo vision, Epipolar geometry, Disparity
TSDFTruncated signed distance function: A volumetric representation storing each voxel's signed distance to the nearest surface. Fusing depth frames into a TSDF is the standard route to real-time reconstruction, with a mesh extracted afterwards.; See also: Voxel, Mesh, Scene reconstruction
Visual odometryVO: Estimating incremental motion from consecutive images. Unlike SLAM it keeps no global map and has no loop closure, so error accumulates without bound. The distinction between the two is a near-universal interview question.; See also: SLAM, Visual-inertial odometry, Loop closure
Visual-inertial odometryVIO: Fusing camera and IMU data to estimate motion. The IMU supplies metric scale and short-term robustness through fast motion or blur, while vision corrects the IMU's rapid drift. The standard approach in headsets and drones.; See also: SLAM, Visual odometry, IMU, Sensor fusion
Voxel: A volumetric pixel — a cell in a regular 3D grid. Voxelising a point cloud imposes regular structure so convolutional methods apply, at the cost of resolution and memory. The basis of many automotive LiDAR detectors.; See also: Point cloud, TSDF, Occupancy grid

Models & Architectures

Anchor box: A predefined box shape a detector refines rather than predicting coordinates from scratch. Anchor design used to require careful tuning to the dataset; anchor-free and DETR-style detectors removed the need.; See also: Object detection, Non-maximum suppression
Approximate nearest neighbour searchANN · Vector search: Finding similar embeddings quickly without comparing against every entry, via indexes such as FAISS or HNSW. What makes visual search over millions of items feasible, and the reason CV roles in retail often touch search infrastructure.; See also: Embedding, Metric learning
AttentionSelf-attention: A mechanism where each element weights every other element by learned relevance. It gives global context in one layer, at a cost quadratic in sequence length — which is why windowed and hierarchical variants exist for high-resolution images.; See also: Vision transformer, Patch embedding
Backbone, neck & head: The standard decomposition of a vision model: a backbone extracting general features, a neck aggregating them across scales, and a task-specific head producing boxes, masks or labels. Backbones are usually pre-trained and reused.; See also: CNN, Feature pyramid network, Transfer learning
CNNConvolutional neural network · ConvNet: A network built from learned convolutional filters, exploiting the locality and translation-equivariance of images. Dominant from 2012 until transformers arrived, and still preferred where data is limited or compute is constrained.; See also: Convolution, Backbone, neck & head, Vision transformer
Diffusion model: A generative model that learns to reverse a gradual noising process, producing images by iterative denoising. Better sample quality and training stability than GANs, at the cost of multi-step inference.; See also: GAN, Synthetic data
EmbeddingFeature vector · Latent representation: A fixed-length vector representing an image or region, positioned so that distance reflects similarity. Production visual search stores these in an approximate nearest-neighbour index rather than comparing exhaustively.; See also: Metric learning, Approximate nearest neighbour search
Feature pyramid networkFPN: A neck architecture combining features across resolutions so a detector can find both small and large objects. Multi-scale handling is one of the practical differences between a detector that works and one that misses small targets.; See also: Backbone, neck & head, Object detection
Foundation model: A large model pre-trained on broad data and adapted to many downstream tasks. In vision this covers CLIP, DINOv2 and SAM — the practical effect is that most teams now fine-tune rather than train from scratch.; See also: Vision-language model, Transfer learning, Self-supervised learning
GANGenerative adversarial network: Two networks trained against each other, a generator producing images and a discriminator judging them. Largely displaced by diffusion on quality and stability, but still used where single-step inference speed matters.; Generative model jobs →See also: Diffusion model, FID
KeypointLandmark: A specific labelled location — a joint, a facial landmark, a corner of an object. Pose estimation predicts sets of keypoints, and evaluation typically uses a distance threshold normalised by object scale.; Pose estimation jobs →See also: 6-DoF pose, Features & descriptors
Metric learningEmbedding learning · Siamese network: Training a model to produce embeddings where similar items are close together, rather than classifying into fixed categories. The right choice whenever the class set changes constantly — retail catalogues, face galleries, defect types.; Metric learning jobs →See also: Embedding, Re-identification
Non-maximum suppressionNMS: Removing duplicate detections by keeping the highest-scoring box and discarding overlapping ones above an IoU threshold. A frequent source of production bugs — too aggressive and you drop genuinely adjacent objects.; See also: Object detection, IoU
Object detection: Locating and classifying multiple objects, returning a bounding box and label for each. The most widely deployed computer vision capability in industry and the entry point for most applied work.; Object detection jobs →See also: Non-maximum suppression, Anchor box, mAP, IoU
Object trackingMOT · Multi-object tracking: Following objects across frames while maintaining consistent identities. Tracking-by-detection dominates: detect every frame, then associate across time using motion and appearance. Identity switches are the characteristic failure.; Visual tracking jobs →See also: MOTA, IDF1 & HOTA, Kalman filter, Re-identification
Optical flow: A dense field of per-pixel motion between consecutive frames. Breaks down on large displacements, occlusion boundaries, textureless regions and reflective surfaces — which is where most real engineering effort goes.; Optical flow jobs →See also: Object tracking, Depth estimation
Patch embedding: Splitting an image into fixed-size patches and projecting each to a vector, so a transformer can treat them as tokens. Patch size sets the trade-off between spatial detail and sequence length.; See also: Vision transformer, Attention
Re-identificationReID: Recognising the same person, vehicle or object across non-overlapping cameras or after a long gap. Built on metric learning rather than classification, since the set of identities is open-ended.; See also: Metric learning, Object tracking
ResNetResidual network: A CNN using skip connections so gradients flow through very deep networks. The residual connection is now near-universal across architectures, and ResNet remains a common baseline backbone.; See also: CNN, Backbone, neck & head
Segment AnythingSAM: A promptable segmentation model that produces masks from a point, box or text hint without task-specific training. Its main industrial impact has been on annotation cost, since it makes model-assisted mask labelling practical.; See also: Segmentation, Annotation, Foundation model
SegmentationSemantic · Instance · Panoptic: Labelling every pixel. Semantic segmentation labels by class without separating objects; instance segmentation gives each object its own mask; panoptic combines both, unifying countable objects with amorphous regions such as road and sky.; Segmentation jobs →See also: U-Net, Dice coefficient, Segment Anything
U-Net: An encoder-decoder segmentation architecture with skip connections between matching resolutions, preserving spatial detail that pooling would lose. Still the default in medical imaging and worth knowing thoroughly for those roles.; Segmentation jobs →See also: Segmentation, Backbone, neck & head
Vision transformerViT: Applying self-attention to images by treating patches as tokens, giving a global receptive field from the first layer. Leads at large data and compute scales; CNNs remain stronger with limited data and cheaper at the edge.; Vision transformer jobs →See also: Attention, CNN, Patch embedding
Vision-language modelVLM · CLIP · Multimodal model: A model with a shared representation of images and text, enabling zero-shot classification, natural-language search and visual question answering. Weak at counting, precise localisation and domains far from web imagery.; Vision-language model jobs →See also: Foundation model, Zero-shot & few-shot, Embedding
Zero-shot & few-shot: Performing a task with no task-specific training examples, or only a handful. Vision-language models enable zero-shot classification by comparing image and text embeddings, removing the fixed-label-set constraint.; See also: Vision-language model, Transfer learning

Training & Data

Active learning: Choosing which examples to annotate next, using model uncertainty, diversity or disagreement. Matters because the labelling budget, not the model, is usually the binding constraint in applied work.; Active learning jobs →See also: Annotation, Label noise
AnnotationLabelling · Ground truth: The labels a model trains against. Usually the largest cost and the main quality bottleneck in applied computer vision — mask annotation is far slower than boxes, and expert domains multiply the cost again.; Active learning & labelling jobs →See also: Label noise, Active learning, Segment Anything
Class imbalance: When some classes are far rarer than others — which is normal, since defects, tumours and collisions are rare by definition. Handled with resampling, loss weighting, focal loss, and metrics that do not reward ignoring the rare class.; See also: Focal loss, Precision & recall
Confidence calibration: Whether a model's stated confidence matches its actual accuracy. Modern networks are typically overconfident, which matters enormously in medical and safety contexts where a confidently wrong output is worse than an abstention.; See also: Precision & recall, Operating threshold
Data augmentation: Synthetically expanding a dataset with transformations — flips, crops, colour jitter, blur, mixup. Effective augmentation is domain-specific: a horizontal flip is fine for pedestrians and wrong for text or medical laterality.; See also: Overfitting, Domain shift, Synthetic data
Data leakage: Information from the evaluation set influencing training, producing results that do not survive deployment. In vision it most often arises from near-duplicate frames, or from splitting by image when the correlation is by patient or scene.; See also: Train / validation / test split
Domain adaptation: Techniques for making a model trained in one setting work in another — fine-tuning on a small target sample, adversarial alignment, test-time normalisation updates, or self-training where no target labels exist.; Domain adaptation jobs →See also: Domain shift, Transfer learning
Domain randomisation: Deliberately varying simulation parameters — lighting, textures, physics — so the real world looks like just another variation. The standard mitigation for the sim-to-real gap.; See also: Sim-to-real gap, Synthetic data, Data augmentation
Domain shiftDistribution shift · Covariate shift: When deployment data differs from training data — a new camera, different lighting, another hospital's scanner, a different country's roads. Usually what separates a model that demos well from one that survives production.; Domain adaptation jobs →See also: Domain adaptation, Model drift, Data augmentation
Focal loss: A loss that down-weights well-classified examples so training focuses on hard ones. Introduced for dense detection, where background examples overwhelm foreground, and now standard for imbalanced problems.; See also: Class imbalance, Dice coefficient
Label noise: Incorrect or inconsistent annotations. Pervasive in real datasets and frequently the true ceiling on model performance — auditing labels often yields more than architecture changes, though it is far less glamorous.; See also: Annotation, Active learning
Model driftConcept drift: Degradation of a deployed model as the world changes. Distinguished from covariate shift in that the relationship between input and label itself changes, which requires relabelling rather than just recalibration.; See also: Domain shift, Model monitoring
Operating threshold: The confidence cut-off converting scores into decisions. Choosing it is a business decision about the relative cost of false positives and false negatives, not a technical default — and strong candidates frame it that way.; See also: Precision & recall, Confidence calibration, Non-maximum suppression
Overfitting: Learning the training set rather than the underlying pattern, producing strong training metrics and weak generalisation. Especially easy in computer vision, where datasets are small relative to model capacity.; See also: Data augmentation, Train / validation / test split, Regularisation
RegularisationWeight decay · Dropout: Techniques constraining a model to improve generalisation — weight decay, dropout, early stopping, augmentation. In inverse problems the term also means the prior that makes an ill-posed reconstruction solvable.; See also: Overfitting, Inverse problem
Self-supervised learningSSL · Contrastive learning: Learning representations from unlabelled data via a pretext task — matching augmented views, predicting masked patches. Most valuable where labels need an expert, such as pathology or industrial defect data.; Self-supervised learning jobs →See also: Transfer learning, Foundation model, Data augmentation
Sim-to-real gap: The performance drop when a model or policy trained in simulation meets real hardware and real sensor noise. Closing it is much of the practical work in robotics learning.; NVIDIA Isaac jobs →See also: Synthetic data, Domain randomisation, Domain shift
Synthetic data: Rendered or generated training data. Commercially important for rare cases that are dangerous or impossible to collect — unusual weather, edge-case collisions, uncommon defects — with the appearance gap as the main obstacle.; See also: Sim-to-real gap, Domain randomisation, Diffusion model
Train / validation / test split: Partitioning data to tune on one subset and report on another. Splitting randomly when data is correlated by patient, site or session leaks information and inflates results — one of the most common serious mistakes in applied CV.; See also: Overfitting, Data leakage
Transfer learningFine-tuning: Starting from a model pre-trained on a large dataset and adapting it to your task. The default approach in applied computer vision, and the reason most production work does not require training from scratch.; See also: Backbone, neck & head, Foundation model, Domain shift

Evaluation & Metrics

Ablation study: Removing or changing one component at a time to establish what actually caused a gain. The discipline that separates researchers who know why their model works from those who only know that it does.; See also: Train / validation / test split
Benchmark overfitting: Community-level overfitting from repeatedly tuning against the same public test set. A reason strong benchmark numbers often fail to transfer, and why hiring managers ask about production failures rather than leaderboard scores.; See also: Overfitting, Data leakage
Dice coefficientF1 for segmentation · DSC: Overlap measure for segmentation masks, closely related to IoU but weighting overlap more generously. The dominant metric in medical image segmentation, and also widely used as a loss function.; See also: IoU, Segmentation, Focal loss
F1 score: The harmonic mean of precision and recall, giving a single balanced number. Convenient for comparison but it assumes precision and recall matter equally, which is rarely true in a real product.; See also: Precision & recall
FIDFréchet inception distance: A measure of how closely a set of generated images matches a real distribution in feature space. The standard generative metric, though it is sensitive to sample size and does not capture per-image quality.; See also: GAN, Diffusion model, LPIPS
IoUIntersection over union · Jaccard index: Overlap between predicted and ground-truth regions, computed as intersection area divided by union area. The basic matching criterion for detection and segmentation, and the input to most detection metrics.; See also: mAP, Dice coefficient, Non-maximum suppression
LPIPSPerceptual similarity: A learned perceptual metric comparing deep features rather than pixels. Correlates better with human judgement than PSNR or SSIM and is standard in super-resolution and novel-view synthesis evaluation.; See also: PSNR, SSIM, FID
mAPMean average precision: The standard detection metric — average precision across recall levels, averaged over classes, usually at several IoU thresholds. A single mAP number hides a great deal, so per-class and per-object-size breakdowns matter more in practice.; See also: IoU, Precision & recall, Object detection
MOTA, IDF1 & HOTA: Multi-object tracking metrics. MOTA counts misses, false positives and identity switches; IDF1 emphasises identity consistency; HOTA balances detection and association. They exist because detection accuracy alone says nothing about identity.; Visual tracking jobs →See also: Object tracking, Re-identification
Precision & recall: Precision is the proportion of predictions that are correct; recall is the proportion of true instances found. They trade off against each other, and which matters more is a property of the application, not the model.; See also: F1 score, Operating threshold, mAP
PSNRPeak signal-to-noise ratio: A pixel-level measure of reconstruction fidelity in decibels. Simple and widely reported, but correlates poorly with perceived quality — a blurry image can score well while looking obviously wrong.; See also: SSIM, LPIPS
SSIMStructural similarity: A quality metric comparing local structure, luminance and contrast rather than raw pixel differences. Better aligned with human judgement than PSNR, though still imperfect.; See also: PSNR, LPIPS

Sensors & Capture

Bayer pattern & demosaicing: Most sensors capture one colour per pixel through a mosaic filter; demosaicing interpolates the missing channels. Working from raw Bayer data rather than processed RGB preserves information that helps low-light and reconstruction tasks.; See also: ISP, Colour space, RAW
Dynamic rangeHDR: The ratio between the brightest and darkest values a sensor can capture in one exposure. Critical in automotive, where tunnel exits and oncoming headlights routinely exceed a standard sensor's range.; HDR imaging jobs →See also: Tone mapping, ISP, Event camera
Event cameraNeuromorphic sensor · DVS: A sensor reporting per-pixel brightness changes asynchronously instead of full frames. Microsecond latency, very high dynamic range and low power, but the output stream is unusable by conventional vision algorithms without conversion.; Event camera jobs →See also: Dynamic range, Rolling vs global shutter
Hyperspectral imagingMultispectral: Capturing many narrow wavelength bands per pixel, producing a spectral signature that identifies materials by composition. Sees chemical differences invisible to an ordinary camera, at the cost of dimensionality and scarce labels.; Hyperspectral imaging jobs →See also: Thermal & infrared imaging, Spectral unmixing
IMUInertial measurement unit: Accelerometers and gyroscopes measuring linear acceleration and angular velocity. High rate and robust to visual conditions but drifts quickly, so it is almost always fused with vision or GNSS.; See also: Visual-inertial odometry, Sensor fusion, Kalman filter
ISPImage signal processor: The hardware pipeline turning raw sensor output into a viewable image — black level correction, demosaicing, white balance, tone mapping, denoising, sharpening. Every modern phone camera is far more ISP than lens.; Computational photography jobs →See also: Bayer pattern & demosaicing, RAW, Tone mapping
LiDAR: A sensor measuring distance by timing reflected laser pulses, producing a 3D point cloud. Accurate range independent of lighting, but expensive, degraded by heavy precipitation, and sparse at long range.; Hiring LiDAR perception engineers →See also: Point cloud, Radar, Motion distortion
Motion blur: Smearing from scene or camera movement during exposure. Degrades feature matching and detection, and forces a trade-off against noise since the alternative is shortening exposure and raising gain.; See also: Rolling vs global shutter, Denoising, Event camera
Motion distortion: Skew in a LiDAR scan because the sensor moves while sweeping. Correcting it requires knowing ego-motion during the sweep, and candidates whose experience is benchmark-only usually have no answer for it.; See also: LiDAR, Rolling vs global shutter, Time synchronisation
Radar: A sensor using radio waves, giving direct velocity measurement through the Doppler effect and working through fog, rain and darkness. Poor angular resolution compared with cameras and LiDAR, which is why it is fused rather than used alone.; RF signal processing jobs →See also: LiDAR, Sensor fusion, SAR
RAW: Unprocessed sensor data before demosaicing, white balance and tone mapping. Higher bit depth and linear response make it preferable for computational photography and scientific work, at the cost of size and complexity.; See also: Bayer pattern & demosaicing, ISP
RGB-D cameraDepth camera: A camera producing colour and per-pixel depth, via structured light, active stereo or time-of-flight. Excellent indoors at short range; most struggle in sunlight, which limits outdoor use.; See also: Time of flight, Structured light, Stereo vision
Rolling vs global shutter: Rolling shutter exposes rows sequentially, skewing fast-moving objects and complicating geometry; global shutter exposes all pixels at once. The distinction matters enormously for SLAM and any high-speed application.; See also: Motion blur, Event camera, Time synchronisation
SARSynthetic aperture radar: Building high-resolution radar images by moving an antenna along a path and combining returns coherently. Images through cloud, smoke and darkness, but speckle noise and geometric effects make interpretation a specialist skill.; SAR imaging jobs →See also: Radar, Speckle
Speckle: The grainy interference pattern characteristic of coherent imaging such as SAR and ultrasound. Not conventional noise — it carries information — so it is filtered with dedicated methods rather than generic denoisers.; See also: SAR, Denoising
Spectral unmixing: Decomposing a pixel's spectrum into constituent materials and their proportions, necessary because one pixel usually covers several materials at typical ground resolution. A core hyperspectral task.; See also: Hyperspectral imaging
Structured light: Projecting a known pattern and inferring depth from how it deforms across the scene. Very precise at close range indoors, which is why it dominates 3D scanning and bin-picking, but it fails in bright sunlight.; See also: RGB-D camera, Time of flight, Stereo vision
Thermal & infrared imagingLWIR · MWIR · NIR: Sensing emitted heat rather than reflected visible light, so it works in complete darkness and reveals temperature differences directly. Low resolution, low texture, little public training data, and often subject to export controls.; Infrared imaging jobs →See also: Hyperspectral imaging, Dynamic range
Time of flightToF: Measuring depth by timing emitted light's return. Compact and fast, used in phones and short-range robotics, but subject to multipath artefacts and interference between multiple sensors.; See also: RGB-D camera, LiDAR, Structured light
Time synchronisation: Aligning timestamps across sensors so measurements can be fused correctly. Along with extrinsics, the most common source of real-world perception bugs — and asking about it is a reliable way to identify hands-on experience.; See also: Sensor fusion, Coordinate frame, Motion distortion
Tone mapping: Compressing a high dynamic range image into a displayable range while preserving local contrast. A perceptual problem with no single correct answer, which is why camera makers differentiate on it.; See also: Dynamic range, ISP

Deployment & Inference

ASIC: A chip built for one purpose, giving the best possible performance per watt at high volume. The endpoint for products shipping in millions of units, at the cost of enormous development expense and total inflexibility once fabricated.; ASIC & custom silicon jobs →See also: NPU, FPGA, Edge inference
Dynamic batching: Collecting requests arriving close together into one batch before inference. Usually the single largest GPU utilisation win when moving from a naive serving setup to a real one.; Triton Inference Server jobs →See also: Throughput, GPU utilisation
Edge inference: Running models on or near the device rather than in the cloud. Chosen for latency, bandwidth cost, privacy or connectivity, and it forces model compression because the compute and power budgets are fixed.; Edge & embedded jobs →See also: Quantisation, NVIDIA Jetson, NPU
FPGA: Reconfigurable hardware giving deterministic microsecond latency and pipeline parallelism — processing can begin as pixels arrive rather than after a full frame buffers. Common in radar front ends and high-speed inspection.; FPGA jobs →See also: ASIC, Latency
Functional safetyISO 26262 · ASIL: The automotive standard governing safety-critical development, defining ASIL levels A to D with requirements on process, traceability and failure handling. It is why automotive perception code looks nothing like research code.; Automotive ECU jobs →See also: Redundancy & fallback, Latency
GPU utilisation: The proportion of available GPU capacity actually doing useful work. An idle GPU costs the same as a saturated one, so utilisation is often the dominant lever on inference cost.; See also: Dynamic batching, Inference
Inference: Running a trained model to produce predictions. The cost centre in production — training happens occasionally, inference happens continuously, so optimisation effort usually pays off here rather than in training.; See also: Latency, Throughput, Quantisation
Knowledge distillation: Training a small student model to reproduce a larger teacher's outputs, often recovering more accuracy than training the small model directly. Common in edge deployment alongside quantisation and pruning.; See also: Quantisation, Pruning
Latency: Time from input to result. In robotics and autonomy it is a correctness property rather than a performance nicety — a perfect answer delivered too late is a wrong answer.; See also: Throughput, Inference, Dynamic batching
MLOps: The practice of building and operating the infrastructure around models — training pipelines, versioning, deployment, monitoring and cost. Becomes essential once a team runs more models than it can manage by hand.; Hiring MLOps engineers →See also: Model monitoring, Reproducibility
Model monitoring: Instrumentation that detects a deployed model degrading — input distribution tracking, confidence distributions, sampled human review. Without it the usual way you learn about failure is a customer complaint.; See also: Model drift, Domain shift
NPUNeural accelerator · Neural Engine: Dedicated silicon for neural network inference, found in phones, cameras and edge SoCs. Far more efficient than CPU or GPU for supported operations — but unsupported layers fall back and destroy the benefit, so architecture choice matters.; See also: Edge inference, Quantisation, ASIC
NVIDIA Jetson: NVIDIA's embedded modules bringing CUDA-capable GPUs to robots, drones and smart cameras. The default platform for on-device inference in robotics because it runs the same stack as the development machine.; Jetson jobs →See also: Edge inference, TensorRT
ONNX: An open interchange format for models, letting a network trained in one framework run on another runtime. The usual bridge between PyTorch training and TensorRT, OpenVINO or ONNX Runtime deployment.; ONNX jobs →See also: TensorRT, Inference
Pruning: Removing weights or channels to shrink a model. Structured pruning removes whole channels and gives real speedups on standard hardware; unstructured sparsity needs specialised kernels to translate into anything faster.; See also: Quantisation, Knowledge distillation
QuantisationINT8 · PTQ · QAT: Representing weights and activations at lower precision, typically INT8, for speed and memory. Post-training quantisation is fast but can cost accuracy; quantisation-aware training recovers most of it at the cost of a retraining cycle.; Model compression jobs →See also: Pruning, Knowledge distillation, TensorRT
Redundancy & fallback: Designing so no single sensor or model failure causes an unsafe outcome — diverse sensing, runtime monitors and degraded operating modes. The practical answer to the fact that no neural network can be proven correct.; See also: Functional safety, Sensor fusion
Reproducibility: Being able to recreate a result — same data version, same code, same hyperparameters, same seed. In regulated domains it is a compliance requirement rather than good practice, which is why model registries exist.; MLflow jobs →See also: MLOps, Ablation study
TensorRT: NVIDIA's inference optimiser and runtime, applying layer fusion, precision calibration and kernel auto-tuning. The standard final step before deploying a vision model on NVIDIA hardware.; TensorRT jobs →See also: Quantisation, ONNX, Edge inference
ThroughputFPS: Items processed per unit time. Distinct from latency: batching raises throughput while raising per-item latency, which is the right trade for backend serving and the wrong one for a control loop.; See also: Latency, Dynamic batching

Datasets & Formats

COCO: A large detection, segmentation and captioning dataset whose JSON annotation format became a de facto standard. Most tooling reads COCO format, so it is usually the path of least resistance for a custom dataset.; See also: Pascal VOC, YOLO format, Annotation
Denoising: Recovering a clean image from a corrupted one. Self-supervised methods such as Noise2Void learn from noisy data alone, which matters in medical and scientific imaging where a clean reference does not exist.; Denoising jobs →See also: PSNR, Speckle, Inverse problem
DICOM: The clinical imaging standard, covering both file format and network protocol, carrying extensive patient and acquisition metadata alongside pixel data. Handling it correctly — including spatial metadata — is a basic requirement in medical imaging roles.; SimpleITK jobs →See also: NIfTI, Registration
ImageNet: The large classification dataset that catalysed modern deep learning in vision. Still the source of most pre-trained backbone weights, though self-supervised alternatives such as DINOv2 increasingly replace it.; See also: Transfer learning, Backbone, neck & head
Inverse problem: Recovering an unknown cause from observed effects — an image from blurred, noisy or partial measurements. Typically ill-posed, meaning small measurement errors produce wildly different results, so the regulariser does most of the work.; Inverse problem jobs →See also: Regularisation, Denoising, Tomographic reconstruction
KITTI & nuScenes: Autonomous driving benchmarks providing synchronised camera, LiDAR and GPS data with 3D annotations. nuScenes is larger and evaluates translation, scale, orientation and velocity error separately rather than with a single IoU.; See also: LiDAR, Bird's-eye view, mAP
NIfTI: A volumetric imaging format common in neuroimaging research, simpler than DICOM while retaining the spatial information needed for measurements to be physically meaningful.; See also: DICOM
Pascal VOC: An early detection and segmentation benchmark whose XML annotation format is still supported widely. Largely superseded by COCO but frequently encountered in older codebases and tooling.; See also: COCO, YOLO format
ROS bag: A recording of all messages on a robot's topics, replayable offline. The primary debugging artefact in robotics — most field failures are diagnosed by replaying a bag rather than by reproducing the situation.; ROS 2 jobs →See also: Coordinate frame, Time synchronisation
Tomographic reconstruction: Building cross-sectional images from projections taken at many angles — the mathematics turning raw CT, MRI and PET acquisitions into diagnostic images. Filtered back-projection and iterative reconstruction are the classical pillars.; Tomographic reconstruction jobs →See also: Inverse problem, DICOM
YOLO format: A plain-text annotation format with one line per object holding class and normalised centre, width and height. Simple and compact, which is why it is common in practical detection pipelines.; See also: COCO, Object detection

Common questions

What is the difference between computer vision and machine vision?

Machine vision is computer vision applied to industrial automation — inspection and measurement on production lines, with controlled lighting and fixed cameras. Computer vision is the broader field, covering everything from medical imaging to autonomous driving.

What is the difference between SLAM and visual odometry?

Visual odometry estimates incremental motion and accumulates drift without bound. SLAM adds a map, loop closure and global optimisation, so revisiting a known place corrects the accumulated error. It is one of the most common computer vision interview questions.

What does IoU mean?

Intersection over union — the overlap between a predicted region and the ground truth, computed as intersection area divided by union area. It is the basic matching criterion for object detection and segmentation, and feeds into metrics such as mAP.

Do I need to know all of these terms to work in computer vision?

No. Most engineers work deeply in one area and know the rest by name only. A perception engineer lives in geometry and sensor fusion; a retail vision engineer lives in detection, embeddings and serving infrastructure. Use the categories to find the cluster that matches the roles you are targeting.

Where to go next

Browse roles by the skills defined here, or — if you are hiring — start from the role guides.