Course Modules

Frames computer vision as an ambiguous inverse problem with a wide application footprint, walks the task family and the end-to-end pipeline, then builds the image-signal foundations (sampling, color, convolution, gradients, scale) and the linear-algebra, calculus, and probability bridge used for the rest of the course.

Derives corner detection from the windowed SSD through the structure tensor and its eigenvalues to Harris and Shi-Tomasi, covers scale-space detectors (FAST, DoG/SIFT, SURF) and their repeatability and invariance, then builds SIFT, ORB, and HOG descriptors with the distances, ratio and cross-check tests, and failure analysis that make matching robust. Quiz 1 falls in this module.

Builds the geometry core: the 2D/3D transformation hierarchy with composition and inversion, homogeneous and projective geometry with incidence and points at infinity, the pinhole camera and its K/R/t projection with scale ambiguity, camera calibration by DLT and Zhang's method, and homography estimation with Hartley normalization. The midterm examination follows this module.

Robust estimation with RANSAC, then two-view geometry through the fundamental and essential matrices and relative pose, stereo rectification and triangulation with depth uncertainty, incremental structure from motion with PnP and bundle adjustment, and finally optical flow and motion analysis in video.

Multi-object tracking with the Kalman filter, gated assignment, and the SORT family; classical bag-of-features recognition with leakage-safe evaluation; the learning-based vision foundations of convolution arithmetic, backprop, losses, and transfer learning; deep object detection across two- and one-stage models; and segmentation and pose estimation with their losses and metrics.