Computer Vision: From Pixels to Convolutions

A build-it-yourself vision course: no OpenCV, no library calls. In every session an AI
mentor walks you through hand-building computer vision's classic weapons in pure numpy.
Starting from "an image is just an array", you'll hand-write 2D convolution and light up
edges with your own Sobel kernel; build histogram equalization, affine warps and Gaussian
pyramids; implement simplified Canny and Harris so the machine finds edges and corners
itself, then match two images with your own descriptors; upgrade hand-made filters into
conv layers, account for parameters and receptive fields, assemble a tiny CNN — and
finally derive conv backprop by hand and train a classifier that truly reads digits.
You'll see it with your own eyes: the kernels the first layer learns look just like the
edge filters you wrote in session one.