Computer Vision Visuals

How one hole turns 3D depth into flat 2D size, and why size alone lies.

Follow pixel differences through Sobel, gradient direction, thinning and connected edges, with a worked 3×3 patch.

How two views shrink your match hunt to one line.

How greedy box selection removes duplicate detections, with exact overlap arithmetic and the cases it gets wrong.

How image motion becomes a vector field, how local gradients constrain it, and why edges, occlusion, and lighting can mislead.

Follow spatial scales, skip connections, channel mixing, and pixel predictions.

Follow scene projection into spatial samples, RGB channels and HSV coordinates, then see what clipping permanently loses.

Move a kernel over actual pixel values, check the weighted sum, and separate kernel orientation, padding and blur trade-offs.

Count eight pixels, accumulate their distribution and build an equalization lookup, then see how the same mechanism exaggerates noise.

Trace a binary structuring element through shrinking, growth, speck removal and hole filling, including the loss of thin structures.

Follow local image changes into a structure tensor, compare Harris and Shi-Tomasi scores, and select spaced corners without confusing detection with matching.

Describe local patches with SIFT gradient histograms or ORB binary comparisons, align their scale and orientation, and see why repeated patterns remain ambiguous.

Turn descriptor neighbors into candidate correspondences, reject ambiguity, fit geometric hypotheses, count inliers and refit without mistaking consensus for truth.

Map a plane between images, divide homogeneous coordinates, sample a warped raster and combine overlapping views while recognizing exposure seams and parallax.

Use measured board corners to separate camera intrinsics, view poses and lens distortion, refine reprojection errors, then examine an underconstrained calibration.

Rectify two views, compare candidate patches, convert horizontal disparity to metric optical depth, and examine consistency, baseline trade-offs and distant-point uncertainty.

Recover sparse scene points and camera poses from overlapping photos, then see why monocular scale and pure rotation leave important ambiguities.

Turn edge pixels into votes for line parameters or circle centers, inspect accumulator peaks, and see how discretization and accidental alignment mislead.

Follow image neighborhoods through learned filters, shared weights, channels, ReLU, stride, pooling and spatial averaging, then expose shift sensitivity.

Compare the mechanisms behind LeNet, AlexNet, VGG, Inception and ResNet, from spatial reduction and activation to parallel scales and residual bottlenecks.

Decode grid-relative centers and anchor sizes into instance boxes, inspect overlap and class scores, then see why suppression can lose crowded objects.

Measure box overlap, match ranked detections, construct precision and recall, and distinguish interpolated AP from averages across classes and IoU thresholds.

Give each object its own mask by retaining proposals, sampling fractional region features with RoIAlign, selecting a class mask and placing it into the image.

Cut an image into patches, project them into tokens, add position, compute attention weights and value blends, and trace the quadratic cost of finer patches.

Connect detections across frames using predicted box motion, Kalman corrections, overlap-based assignment and explicit track lifetimes, then expose identity loss.

Trace a four-pixel image through Gaussian corruption, noise-prediction training, clean estimates and stochastic reverse sampling, ending with an ambiguous inverse.