140 likes | 304 Views
CSE 140: Computer Vision. Camillo J. Taylor Assistant Professor CIS Dept, UPenn. What is Computer Vision. Recover information about the scene from one or more images Relevant questions What’s out there Where is it Where is it going. Computer Vision as an Inverse Problem.
E N D
CSE 140: Computer Vision Camillo J. Taylor Assistant Professor CIS Dept, UPenn
What is Computer Vision • Recover information about the scene from one or more images • Relevant questions • What’s out there • Where is it • Where is it going
Computer Vision as an Inverse Problem • Given information about the geometry of the scene, surface reflectance properties, lighting configuration and camera model we can generate an image • The problem of computer vision is to infer these properties of the scene from the image data (a 2D array of numbers]
Outline • Reconstruction • Recognition
Stereo • Given two images of a scene taken from known positions with calibrated cameras recover the 3D structure of the environment.
Correspondence Problem • Stereo algorithms attempt to automatically detect correspondences between points in the two images. • Once these correspondences have been determined, the locations of these points in space can be determined from a simple triangulation
+ D å = + + - + 2 SSD ( i , d ) ( I ( i j d ) I ( i j )) r l = - D j SSD matching • The sum of squared differences is a commonly used matching criterion in stereo algorithms
Constraints that can be employed in matching • Uniqueness • Ordering • Smoothness
Recovering 3D models from 2D image • For single images we can only recover the geometry if we make some assumptions about the shape of objects that we’re looking at
Recognition • One statement of the recognition problem • Given a 3D model of a target object determine whether that object appears in a given image • In the alignment approach to recognition the computer hypothesizes matches between features in the model and features in the image and then tests whether any other features support this match
Other approaches to recognition • Image based • Directly compare stored image to acquired image to determine whether or not they are similar • Useful in face recognition
Recognizing articulated objects • David Forsyth’s Horse finder