Method for generating a low-dimensional representation of high-dimensional data

• US 20060045353A1
• Filed: 09/02/2004
• Published: 03/02/2006
• Est. Priority Date: 09/02/2004
• Status: Active Grant

• Current Assignee: Mitsubishi Electric Research Laboratories, Inc. (Mitsubishi Electric Corporation)
• Original Assignee: Mitsubishi Electric Research Laboratories, Inc. (Mitsubishi Electric Corporation)
• Inventors: Brand, Matthew E.
• Granted Patent: US 7,412,098 B2
• Time in Patent Office: Days
• US Class Current: 382/224
• CPC Class Codes: G06F 18/213 Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods

Abstract

A method represents a class of objects. A set of samples for the objects in the class is acquired, there being one sample for each object, and each sample includes a plurality of data values representing characteristics of the object. The samples are grouped into subsets such that each subset intersects at least one other subset. For each subset, a low-dimensional parameterization is determined. Nullspaces of the low-dimensional parameterizations are averaged to obtain a matrix whose nullspace contains a low-dimensional representation of the class of objects.

Claims

1. A method for representing a class of objects, comprising:

acquiring a set of samples for the objects in the class, there being one sample for each object, and each sample includes a plurality of data values representing characteristics of the object;

grouping the samples into subsets such that each subset intersects at least one other subset;

determining, for each subset, a low-dimensional subset parameterization; and

averaging nullspaces of the low-dimensional subset parameterizations to obtain a matrix having a nullspace containing a low-dimensional representation of the class of objects.

2. The method of claim 1, further comprising:

reducing noise in the set of samples by determining a difference between the set of samples and the nullspace containing the low-dimensional representation of the class of objects.

3. The method of claim 1, further comprising:

mapping a column nullspace of the matrix to the subset parameterizations to generate the low-dimensional representation of the class of objects.

4. The method of claim 1, further comprising:

acquiring a next set of samples for the objects in the class;

grouping the next samples into next subsets;

determining, for each next subset, a low-dimensional subset parameterization;

averaging nullspaces of the low-dimensional subset parameterizations of each next subset and each subset intersecting a next subset, to obtain a filter; and

filtering the set of samples and the next set of samples to update the representation of the class of objects.

5. The method of claim 1, in which the samples are images of human faces.

6. The method of claim 1, in which the samples are models of human faces.

7. The method of claim 1, further comprising:

rendering the representation on an output device.