## CS9629

tentative list of topics

This course presents many standard computer vision problems and their solution methods using common algorithms (e.g. dynamic programming, shortest paths, graph cuts, minimum ratio cycle). The studied image analysis problems provide an intuitive visual environment helping better understanding of such optimization methods. A tentative list of topics is given below.
- Overview of Imaging Modalities: photo (camera model), video, medical, etc.
- Elements of Image Pre-Processing (gamma and window/center correction, histogram equalization)
- Features and Filtering: colors, edges, corners, SIFT, etc.
- Basic Image Segmentation

- unsupervised and supervised segmentation (e.g. background subtraction vs. photoshop)

- "naive" methods: thresholding, region growing, watersheds

- clustering techniques: mean-shift, K-means (variance clustering criteria)

- boundary regularization: livewire, active contours (snakes), graph cuts, etc.

- combining (known) color model and boundary smoothness: objective function (energy), graph cuts

- Model Fitting

- probability models: Gaussians, GMM, histograms, Parzen density

- color model fitting in segmentation: Chan-Vese, Zhu-Yuille, grabcut. From variance to entropy-based color clustering.

- geometric models (lines, planes, homographies) for reconstruction (structure, panoramas, pose)

- inference methods: ML, least-squares, outlier robustness, L2 vs L1 metric, RANSAC

- multi-model fitting: hard assignemnts (K-means) vs. soft assignments (EM)

- General Image Labeling and Optimization

- applications: restoration, stitching, detection, segmentation, motion, etc.

- inference models: Markov Random Fields (MRF), Minimum-Description Length (MDL), sparcity

- optimization methods: ICM, annealing, binary and multi-label graph cuts (submodularity, a-expansion)

- pair-wise and high-order regularization: smoothness, boundary length and curvature, color-consistency, label cost, etc.
- Correspondence, Matching, and Reconstruction

- template matching, image registration, tracking, motion layers

- shape registration, pictorial structures

- stereo, volumetric reconstruction, laser scanning, kinect

- structure from motion

- HW1: color quantization and superpixes (RGB and RGBXY clustering) via K-means and mean/medoid-shift (extra credit)
- HW2: interactive segmentation interfaces: snakes (edges), graph cuts (seeds+edges+color model), grabcut (box+edges+model fitting - extra credit)
- HW3: line fitting (RANSAC, multiple lines), real image example with feature detector (extra credit)
- mini project for grad students on stereo: windows, scan-line, TV, Potts (3 out of 4)