Computer vision mini projects.

Determining the scene an image was taken in from a bag of visual words. A plain histogram throws away all spatial information, so spatial pyramid matching adds progressively finer grid subdivisions to capture where features appear, not just which ones.

Matching points with a BRIEF descriptor, estimating the homography between two views of the same plane, and warping a new cover onto a textbook. Extended to a full AR application overlaying video frame by frame.

Tracking via optical flow: a template tracker with drift correction, then dominant affine motion estimation that finds moving objects by modelling the background and treating whatever disagrees as the target, plus the inverse-compositional extension for efficiency.

From two images to a metric point cloud: the fundamental matrix via the eight-point algorithm, the essential matrix from calibrated intrinsics, and triangulation, with automatic matching along epipolar lines.

Reading handwritten text out of an image, which is as much classical image processing as learning: the characters have to be found, boxed, grouped by line and sorted before the network sees anything. Plus autoencoders, and a PyTorch CNN at 90.2% against the fully connected network's 76%.

Shape from shading under a set of lighting directions, first calibrated to recover albedo and normals, then uncalibrated, where the bas-relief ambiguity means the images genuinely cannot distinguish between a whole family of shapes.