25553: Computer Vision
Course Name: Computer Vision
Course Number: 25553
Prerequisite(s): 25765 (Digital Signal Processing)
Co-requisite(s): -
Units: 3
Level: Postgraduate
Last Revision: Fall 2014

Description: 
The ultimate goal in computer vision is to emulate human vision, including making inferences based on visual inputs. In this course, the students will learn fundamental approaches in computer vision.
 
Syllabus:
  • Filtering: smoothing, removing noise, convolution, image derivatives
  • Edge detection: gradient operators (Prewitt and Sobel), Marr-Hildreth edge detector (Laplacian of Gaussian), Canny (gradient of Gaussian)
  • Interest point detection: local features, interest points, Harris corner detector
  • Scale-invariant feature transform: extracting key points, scale space, outlier rejection, orientation assignment, descriptors, key point matching
  • Optical flow: brightness constancy equation, smoothness constraint, Horn-Schunck algorithm, Locus-Kanade algorithm
  • Pyramids: Gaussian and Laplacian pyramids, optical flow with pyramids for large motion
  • Motion model: 3D rigid motion, rotation using Euler angles, orthographic and perspective projections, affine transformation, displacement model, instantaneous model, homography
  • Global motion: Bergen et al. method, coarse-to-fine global flow estimation, image warping, Mann and Picard method, generating mosaic
  • Camera model: camera calibration, camera model, finding camera location, camera orientation, camera parameters
  • Fundamental matrix: RANSAC, derivation of Fundamental matrix, essential matrix, normalized 8-point algorithm, robust fundamental matrix estimation
  • Mean-shift tracking: mean-shift theory and its applications, mean-shift vector, parametric and non-parametric density estimation, kernel density estimation, real modality analysis, likelihood maximization using mean-shift
  • Kanade-Locus-Tomasi tracker: KLT tracking algorithm, finding alignment, KLT-Baker algorithm
  • Structure from motion: Tomasi and Kanade factorization method
  • Hough transform: line fitting (Cartesian-polar form), circle fitting, generalized Hough transform for arbitrary shape, r-table, rotation, and scale-invariant shape fitting
  • Bag-of-features: image classification by histogram of features
  • Face recognition: eigenface method, within-class, and between-class variation, Fisherfaces
  • Stereo: shape from a stereo, rectification, correspondence search, correlation measures, block matching, simulated annealing, Barnard stereo method

References:
  • R. Szeliski, Computer Vision: Algorithms and Application
  • D. A. Forsyth, J. Ponce, Computer Vision - A Modern Approach

 
Last Update: 2024-07-11