1st Day Lecture 1: Intro
Goal of Vision To understand and interpret the image. Images consist of many different patterns – grass, faces, crowds.
Vision is easy for Humans Because a very large part of our brain does vision. Half the Cortex does Vision. (Much more than does mathematics or computer science or other ‘high level’ tasks.)
Vision is very difficult Because images are complex and ambiguous. Left panels (top) shows two bicycles Left panels (bottom) show intensity plots I(i,j).
Vision as Decoding Vision is an Inverse Inference Problem
Bayes Theorem. Bayes (left) uses prior knowledge to resolve ambiguity (right).
Lecture 2: Images and Filters
Statistics of Image Gradients Left: Everywhere. Right: On and Off Edges.
Lecture 3: Edges
Edges: Sowerby Dataset Top: Example of Images and groundtruth Bottom: Canny (left), Statistical method (center). Loglikelihood ratios (right)
Edges: Combing Scales Results: Chernoff performance measures (risk). Triangles (grad I). Cross (Harris-Nitzberg).
Edges: Combining Directions Results using combinations of oriented filters (Gabors).
Edge Detection is Hard Distributions overlap: ROC curves.
Lecture 4: Weak Smoothness
Steepest Descent and Variations Steepest Descent and Discrete Iterative.
Lecture 5 (Manhattan World)
Manhattan world The world has Manhattan structure. And humans make mistakes when it does not. Identical twins in Ames room.
Geometry Projection and Vanishing Points
Manhattan Results Good results for Indoor Images
Manhattan Images Results for City Scenes
Manhattan Countryside Some images in the country also have Manhattan structure.
Non-Manhattan Images Some images are not Manhattan – verify by model selection: compare P_{man} to P_{null}
Lecture 6 Image Classification – independent.
Sowerby and San Francisco Examples.
Results: Color only (left). Texture only (right). Sowerby only.
Results: Color and Texture: Sowery (left), San Francisco (Right).
Examples. Sowerby:
Examples San Francisco
Medical Applications Apply similar ideas to medical images. Tumor detection.