Stanford OnlineStanford AA274A Principles of Robotic Autonomy | Autumn 2019 | Image Processing
Stanford Online
Vision tasks and features 0:05
The lecture shifts from 3D reconstruction to the broader use of vision for semantic reasoning. Images are first treated as pixel grids, but the goal is to work with features such as corners, edges, and blobs. These key points let you match points of interest across images and support tasks like stereo correspondences, structure from motion, and object detection. A butterfly image shows a detector picking out salient points that later feed higher-level understanding.
Filtering and correlation 6:03
An image processing pipeline has four steps: filtering, feature detection, feature description, and higher-level processing. Today covers the first three. Filtering cleans an image, often to remove noise before derivatives are taken, because derivatives amplify noise. In the spatial domain, a filter uses a neighborhood, such as a 3x3 window, and an operation on those pixels. Linear filters are written as correlation with a mask or kernel, such as an all-ones average filter.
Smoothing filters 21:02
A moving average filter replaces each pixel with the weighted average of the pixels under the mask. The mask is usually normalized so a flat image keeps the same intensity. A Gaussian smoothing mask does the same job with more weight at the center and less at the edges. It is the usual first step before taking derivatives, because derivatives can amplify noise.
Convolution and derivatives 25:01
Convolution is like correlation, but the mask is flipped. That flip gives associativity, so you can combine masks first and save work. For discrete images, the derivative is defined with finite differences, not calculus limits. A simple two-sided difference uses values on both sides of a pixel, while Sobel masks use neighboring rows and columns to make the estimate more robust.
Edge detection basics 42:00
An edge is a sharp change in intensity, like white to black. The method first smooths the image to remove noise, then applies a derivative filter and looks for large responses. Good edge detection needs accuracy, good location, and a single response, so one real edge does not turn into several thin ones.
Corners and Harris 55:30
After thresholding, the result is cleaned with non maxima suppression so the lines stay thin. Corners are then introduced as points where intensity changes in all directions. A window is moved around the image and its response is tested. The Harris detector turns this idea into derivatives and eigenvalue decomposition. It is rotation invariant and stable to uniform intensity shifts, but it can fail when scale changes.
Descriptors and neural nets 1:02:00
Scale changes can break corner matching and hurt stereo vision and visual odometry, so detectors were extended beyond rotation and intensity changes. Descriptors then record the neighborhood around a key point so features can be compared across images. SIFT is a common detector and descriptor, with a 128-dimensional vector and good scale and rotation behavior. ORB and BRISK are lighter. Neural networks now do much of this work, but they need data and are harder to interpret or trust.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
