Atlas 136
Practical computer vision atlas — algorithms, models, and concepts.
Recently added8
Computes each output element as a learned, input-dependent weighted average of value vectors, letting every element aggregate information fr…
Contrastive image–text pretraining on 400M web pairs: dual encoders in one embedding space make natural language the classifier — zero-shot …
Data-efficient image transformers: ViT's architecture unchanged, made ImageNet-1k-competitive by a heavy augmentation/regularization recipe …
Self-distillation with no labels: an EMA teacher, multi-crop training, and centering+sharpening yield ViT features whose frozen k-NN accurac…
A 6.7B-parameter self-supervised ViT trained on 1.7B curated images, with gram anchoring to stop dense-feature degradation at scale — one fr…
Training a student network to match a teacher's softened output distribution — the temperature softmax, the T² gradient scaling, hard-label …
Standardises intermediate activations over a chosen index set, then restores capacity with a learned affine transform; batch, layer, instanc…
Algorithms51
Solve the hand-eye equation AX=XB jointly for rotation and translation by parametrising rigid motions as unit dual quaternions and extractin…
Recover the constant rigid transform from a robot gripper to a rigidly mounted camera by solving the AX=XB equation in two stages — modified…
Real-time frontal-face detection by sliding a fixed 24×24 sub-window across a grayscale image at multiple scales, scoring each position with…
Partition an image into perceptually coherent regions by a Kruskal-style greedy merge over a pixel graph, accepting an inter-component edge …
Extract a foreground from a colour image using a single bounding rectangle as the only required input by alternating Gaussian mixture compon…
Compute the global minimum of a binary region-and-boundary MRF energy as a single s-t min-cut on a pixel graph; user-marked seeds enter as h…
Founding random-sample-consensus paradigm: fit a parametric model to data containing an unknown fraction of gross outliers by drawing minima…
Robust estimator that eliminates the user-tuned inlier threshold by treating the noise scale σ as a random variable on [0, σ_max] and margin…
Engineering decomposition of practical RANSAC into four pluggable stages — sampling (PROSAC), model verification (SPRT), local optimisation …
Rectifies a monocular image to a metric overhead (bird's-eye) view by constructing the rectifying homography in closed form from two CNN-reg…
Detect a target object class in arbitrary images by scoring every position and scale in a HOG feature pyramid with a mixture of star-structu…
Refine pixel-level chessboard corner positions to sub-pixel accuracy by nonlinear least-squares fitting a seven-parameter ideal blurred-corn…
Detects scale- and rotation-invariant blob keypoints as scale-space maxima of the Hessian determinant, approximated with box filters on an i…
Detect thin step edges in greyscale images by smoothing with a Gaussian, computing gradient magnitude and direction, suppressing non-maxima …
Calibrate any central catadioptric or fisheye camera from a few planar checkerboard views by fitting a radially-symmetric Taylor-polynomial …
Non-iterative O(n) solver for the calibrated Perspective-n-Point problem: express the n reference points as weighted sums of four virtual co…
Gradient-vote operator that highlights pixels of high local radial symmetry — bright/dark blobs and approximately circular features. Each pi…
Affine extension of FRST: each pixel votes along a corrected direction $\hat V = G M G^{-1} M^{-1} \nabla I$ at radius $n$, where $G = R D \…
Detect checkerboard X-corners by computing a four-quadrant corner likelihood at each pixel using axis-aligned and 45°-rotated prototype filt…
Post-process a partially detected checkerboard by training two Gaussian processes (one per pixel coordinate) on the allocated (boardXY, boar…
Recover camera intrinsics from one or more views of one or more planar targets via the same two IAC-on-homography constraints as Zhang's met…
Two-stage 1987 camera calibration that uses the radial alignment constraint to recover extrinsics and image scale linearly from a precision …
Extend Tsai's radial alignment constraint to a non-frontal sensor by modelling lens–sensor tilt as a 2-DoF rotation, projecting observations…
Detect checkerboard X-junctions by approximating a localized Radon transform with 1-D box filters on rotated copies of the image; the per-pi…
Recover camera intrinsics, radial distortion, and per-view extrinsics from at least three images of a planar pattern at different orientatio…
Detect chessboard X-junctions in heavily blurred or high-resolution images by computing a 16-sample circular x-corner intensity at every lev…
Recover the largest visible checkerboard subgraph from a partially occluded pattern by running VF2 subgraph isomorphism against a model grap…
Recover the integer $(i, j)$ grid coordinate of every corner in a checkerboard calibration image by Delaunay-triangulating the corners, merg…
A chessboard-specific corner detector: scores each pixel by how well its local neighborhood matches an alternating bright-dark X-junction pa…
Segment-test corner detector on a 16-pixel Bresenham ring of radius 3 around each candidate; classifies a point as a corner when N contiguou…
Compute a fixed-length descriptor for an image window by binning pixel gradients into 8×8 cells of 9 unsigned-orientation histograms, normal…
1981 closed-form linear method for relative orientation of two viewpoints from eight calibrated point correspondences, introducing the bilin…
Encodes a Gaussian-smoothed image patch around a detected keypoint as a 128/256/512-bit binary string by running a fixed table of pairwise p…
Detects rotation-invariant oriented keypoints by running FAST-9 on a √2 image pyramid, ranking by Harris cornerness, and assigning orientati…
Stitch two-plane outdoor panoramas by clustering SIFT correspondences into a ground group and a distant group via spatial K-means, fitting o…
Stitch two images under moderate parallax by replacing the global affine with a per-feature deviation field, regularised to be smooth via a …
Compute the fundamental matrix from n ≥ 8 point correspondences by conditioning the linear DLT system via a similarity normalisation, recove…
Replace a global homography with a spatially varying field of homographies, each fit by a per-cell weighted DLT (Moving DLT) on the same poi…
Detect a full planar checkerboard in an image by reducing the gradient-magnitude edge set to a single-pixel centreline graph, extracting inn…
Detect every corner of a chessboard calibration pattern and assign it an integer grid coordinate by counting ring-alternations to locate X-j…
Detect and decode a self-identifying checkerboard calibration pattern: saddle-point corners from a Hessian response, grid reconstruction via…
Scores each pixel by the Harris response R = det(M) − k·tr(M)², where M is the gradient covariance matrix summed over a Gaussian window; ret…
Scores each pixel by the smaller eigenvalue of the gradient structure tensor M; returns integer pixel locations where that eigenvalue exceed…
Calibrated Euclidean rectification that builds a new pair of projection matrices sharing a common orientation from the two known PPMs, yield…
Computes a rectifying homography pair from the fundamental matrix alone, sending the epipole to infinity with a quasi-affine perspectivity a…
Computes a rectifying homography pair from a known fundamental matrix by factoring each homography as shearing × similarity × projective, ch…
Rectification by polar reparametrisation around the epipole, requiring only the oriented fundamental matrix and remaining bounded even when …
Optical flow that replaces the quadratic data and smoothness penalties of variational flow with redescending M-estimators, solved by SOR wit…
Dense optical flow recovered by minimising a variational energy that combines the brightness-constancy constraint with a global smoothness p…
Iterative Newton-Raphson method that estimates the parametric warp between two images by linearising the residual and solving the resulting …
Detects keypoints as scale-space extrema in a Difference-of-Gaussian image pyramid, refines location and scale by 3D quadratic interpolation…
Models50
Practical click-based interactive segmentation that runs each click as a small local-crop forward pass (Segmentor on a Target Crop, Refiner …
Feedforward click-based interactive segmentation: HRNet+OCR encoder-decoder taking RGB + positive/negative disk-encoded clicks + previous ma…
Two-stage CNN object detector that replaces external Selective Search / EdgeBoxes proposals with a learned Region Proposal Network sharing c…
Single-stage CNN object detector that frames detection as one regression problem from full-image pixels to a 7×7×30 tensor of grid-cell box …
Dense semantic segmentation by repurposing an ImageNet classifier with atrous (dilated) convolution to preserve spatial resolution, an Atrou…
Two-stage instance segmentation by adding a parallel FCN mask branch to Faster R-CNN — per-class binary masks predicted at each RoI under a …
Symmetric encoder-decoder fully-convolutional network for dense pixel-wise biomedical image segmentation — contracting path with channel-dou…
Encoder-decoder CNN for dense pixel-wise classification — converts ImageNet classifiers into fully convolutional networks via 1×1-conv reint…
A monocular depth network trained for zero-shot cross-dataset transfer by mixing incompatible depth datasets under a scale-and-shift-invaria…
Mobile-CPU CNN image classifier discovered by platform-aware neural architecture search: an RNN controller trained with reinforcement learni…
Efficient mobile CNN backbone built from inverted-residual blocks with a linear bottleneck — depthwise-separable convolution expanded to a w…
Mobile-CPU-latency-targeted CNN backbone found by combined platform-aware NAS and NetAdapt, built from MobileNetV2 inverted-residual blocks …
Light-weight specialist real-time detection transformer that discovers an accuracy-latency Pareto frontier for any target dataset from a sin…
Two-branch (bilateral) CNN for real-time semantic segmentation: a wide shallow path preserves spatial detail while a deep narrow path with g…
Real-time semantic segmentation CNN whose shared shallow 'Learning to Downsample' prefix feeds both a deep low-resolution global-feature bra…
Hierarchical Transformer encoder (MiT) producing multi-scale features at $1/4, 1/8, 1/16, 1/32$ without positional encodings, plus an all-ML…
End-to-end object detector that recasts detection as direct set prediction — CNN backbone (ResNet-50/101) extracts $H/32 \times W/32$ featur…
CNN backbone family for dense prediction that maintains a high-resolution branch throughout the network and runs four parallel multi-resolut…
Masked Autoencoder — self-supervised pretraining for Vision Transformers: randomly mask 75 % of input patches, feed the visible 25 % through…
Universal image segmentation family — MaskFormer (v1, NeurIPS 2021) reframes semantic segmentation as **mask classification**: predict a set…
Promptable segmentation foundation model family — SAM (v1, 2023) introduces image-prompt segmentation with a heavy ViT-H encoder and lightwe…
Vision Transformer — a pure-transformer image classification backbone that treats an image as a sequence of fixed-size patches: split RGB im…
Family of very deep CNN image classifiers (18 to 152 layers) built from residual blocks $y = \mathcal{F}(x, \{W_i\}) + x$ that reformulate e…
Eight-layer convolutional neural network for 1000-class image classification on ImageNet, trained end-to-end on two GPUs with ReLU activatio…
Twenty-two-layer CNN built from Inception modules — parallel 1×1, 3×3, 5×5 convolutions and 3×3 max-pool concatenated along the channel axis…
Family of very deep CNN image classifiers (11 to 19 weight layers) built from stacked 3×3 convolutions with stride 1 and 2×2 max-pool stride…
First learned per-pixel checkerboard X-corner detector: a three-convolutional-layer CNN with 2,939 parameters trained with mean-squared-erro…
Fully-convolutional CNN that jointly detects interest points and computes 256-D descriptors in a single forward pass, trained without human …
Fully convolutional network that regresses a per-pixel checkerboard-corner response map; trained with weighted cross-entropy and paired with…
Lightweight CNN that jointly detects keypoints, extracts 64-D dense descriptors, and refines semi-dense matches from coarse descriptor pairs…
Contrastive image–text pretraining on 400M web pairs: dual encoders in one embedding space make natural language the classifier — zero-shot …
Data-efficient image transformers: ViT's architecture unchanged, made ImageNet-1k-competitive by a heavy augmentation/regularization recipe …
Self-distillation with no labels: an EMA teacher, multi-crop training, and centering+sharpening yield ViT features whose frozen k-NN accurac…
Hierarchical vision transformer with shifted-window attention: linear complexity in image area, CNN-style multi-scale feature maps, and a dr…
Millisecond-latency one-class anomaly detection combining a loss-induced-asymmetric student-teacher branch for structural defects with a fea…
Training-free industrial anomaly detection: a single forward pass over defect-free images builds a coreset-subsampled memory bank of locally…
Pixel-precise anomaly segmentation from an ensemble of student networks trained only on anomaly-free images to regress a fixed off-domain te…
A monocular relative-depth foundation model that scales training to 62M unlabeled images via a teacher pseudo-labeling data engine, a CutMix…
A monocular depth foundation model that trains its teacher purely on synthetic images for label precision, then distills to a student over 6…
A self-supervised ViT trained on a curated 142M-image dataset that yields general-purpose visual features usable frozen — via kNN or linear …
Lightweight SAM family — replaces SAM's heavy ViT-H image encoder (632M params, ~452 ms on a single GPU) with a distilled TinyViT encoder (5…
Three-stage learning-based camera calibration pipeline: a CNN regresses radial-distortion-correction parameters, a UNet predicts per-corner …
A 6.7B-parameter self-supervised ViT trained on 1.7B curated images, with gram anchoring to stop dense-feature degradation at scale — one fr…
A single plain-transformer model that predicts spatially consistent geometry — depth plus camera rays — from one to many images, with or wit…
A feed-forward network that regresses two dense pointmaps in a shared coordinate frame from an uncalibrated, unposed image pair, jointly rec…
A 3D-grounded image matcher that adds a dense local-descriptor head and an InfoNCE matching loss on top of DUSt3R's pointmap regression, wit…
A large feed-forward transformer that predicts cameras, depth maps, point maps and 3D point tracks for one to hundreds of views in a single …
Adaptive-depth Transformer matcher for sparse local features: stacks 9 self+cross-attention layers with rotary positional encoding and a per…
Detector-free dense feature matcher: shared CNN backbone produces coarse and fine feature maps, a Linear Transformer with interleaved self- …
Concepts35
Computes each output element as a learned, input-dependent weighted average of value vectors, letting every element aggregate information fr…
Training a student network to match a teacher's softened output distribution — the temperature softmax, the T² gradient scaling, hard-label …
Standardises intermediate activations over a chosen index set, then restores capacity with a learned affine transform; batch, layer, instanc…
The linear, shift-invariant operation that produces each output pixel as a kernel-weighted sum of input pixels in a local neighbourhood.
The framework that poses image labelling and segmentation as minimising an objective combining a per-pixel data term and a pairwise smoothne…
A discrete multi-resolution representation — a sequence of images at progressively coarser resolution, each smoothed and downsampled from it…
A precomputed prefix-sum array that returns the sum of pixel values over any axis-aligned rectangle in constant time with four array reads.
Reducing a dense response map or a set of overlapping detections to a sparse set of local maxima by discarding every element that is not str…
The projective map from 3-D scene points to 2-D image pixels through a single centre of projection, parameterised by an intrinsic matrix and…
Estimating a geometric entity defined only up to scale by stacking constraints into a homogeneous linear system and taking the smallest righ…
Random sample consensus — a paradigm for fitting a parametric model to data containing an unknown fraction of gross outliers, by drawing min…
Verify candidate calibration-pattern corners by constructing a graph over them (Delaunay triangulation, k-nearest-neighbours, or proximity) …
Mechanisms that inject token order into permutation-invariant attention: absolute encodings added to embeddings, learned tables, and rotary …
Survey of label-free visual representation learning: contrastive objectives (SimCLR, MoCo) and negative-free self-distillation (BYOL, DINO),…
Predicting per-pixel scene depth from a single image — the scale ambiguity that forces relative (affine-invariant) versus metric formulation…
Joint nonlinear least-squares refinement of all camera parameters — and, in structure-from-motion, all 3-D points — that minimises the total…
A feed-forward network that builds a spatial hierarchy of learned features by alternating weight-shared convolution layers, pointwise nonlin…
Recovery of the 6-DOF rigid transformation — rotation and translation — relating a camera to a scene, an object, or a second camera.
Mathematical models for departures from the ideal pinhole projection — radial barrel/pincushion, tangential decentering, thin-prism — and th…
A two-line similarity transform — translate the point centroid to the origin, isotropically scale so the average distance is √2 — that condi…
An invertible projective transformation of the plane, represented by a 3×3 matrix defined up to a non-zero scalar, mapping points between tw…
The 2-vector of partial derivatives of image intensity with respect to spatial coordinates, measuring the rate and direction of brightness c…
A one-parameter family of images obtained by progressively blurring an input image with Gaussians of increasing standard deviation, providin…
Sequence-to-sequence architecture assembled entirely from attention and position-wise feedforward sublayers, each wrapped in a residual conn…
One-class detection and localisation of defects from defect-free training images only, organised by the structural-versus-logical anomaly di…
Fixed-length vectors encoding the local image appearance around a keypoint, built so the same physical point yields similar descriptors acro…
Twenty-five years of methods for finding the inner corners of a planar checkerboard calibration target — from Harris-on-thresholded-images t…
A scalar response computed from the determinant of the image Hessian, negative at saddle points (X-corners) and zero at flat regions, edges,…
A 2011–2013 lineage of stitching methods that replace the single global homography with a spatially varying warp field — fitted as either tw…
The intrinsic projective geometry of two views of a scene, encoding the constraint that a point visible in one image must lie on a specific …
A symmetric 2×2 matrix formed by summing the outer products of the image gradient over a local window, encoding the dominant orientation and…
Warping a stereo pair so corresponding epipolar lines become collinear image rows, reducing dense correspondence to a scanline search; calib…
Recovering 3D geometry — point maps, depth, and camera poses — directly from images in a single network pass, replacing the detect-match-tri…
Establishing keypoint correspondences between two images by comparing descriptors and resolving them into a consistent partial assignment — …
The apparent 2-D velocity field of image brightness between consecutive frames, recovered from the spatio-temporal gradient under the bright…