Image Similarity Based Image Analysis
we consider the analysis of thousands of unorganized,
low resolution images of an object. With very low resolution images,
standard computer vision techniques of finding corresponding points
and solving for image warping parameters or 3D geometry may fail. Two
recent techniques in statistical pattern recognition, locally linear
embedding (LLE) and Isomap, give a mechanism for finding the structure
underlying point sets for which comparisons or distances are only
meaningful between nearby points.
We explore these methods to
simultaneously compute camera position and object pose for thousands
of images using nothing but a global similarity measure between
images.
We also have a method for computing the MDS solution to embed points
on a sphere.
Metric Pose Estimation from Similarity Measures
Randomly Arranged Images (algorithm input)
Images Arranged by Isomap
Images Arranged by Isomap, Linear Transform, and Smoothing
Using Thousands of Images of an Object
Robert Pless and Ian Simon
Computer Vision, Pattern Recognition and Image Processing,2002
(PDF, 330 kb)