Skip to content
anythingai.space
AI Foundationsbeginner13 min · 9 steps

Vision Models: CNNs, Classification and Detection

How convolutional networks turn pixels into labels, and how detectors from R-CNN to YOLO find every object in a frame.

pixelsSame cat, new position or lighting → everyrule has to changePhoto224 × 224 pixelsHand-written rulesSIFT · HOG features“cat”?
dataerrorresponseheavy work
1 / 9
ProblemStep 1 of 9

To a computer, a photo is a grid of numbers

A phone photo is just brightness values in a grid. A cat can sit anywhere in that grid, at any size, lit from any side. Hand-written rules (“pointy ears are here”) break immediately.

Older pipelines used hand-crafted features such as SIFT and HOG. They were interpretable, but nowhere near as powerful as features a network learns by itself.

InsightThe breakthrough was letting the network learn its own features from millions of labelled photos.

Lesson 5 of 34 in AI Foundations