Invariant and Calibration-Free Methods in Scene Reconstruction and Object Recognition.

Abstract

This is the final technical report for a DARPA project (#MDA 972-91-C-0053) on object recognition through the use of Geometric Invariants. It gives an extended account of the work that was done at GE-CRD and by certain collaborating researchers during the time of this project. The report considers the subjects of reconstruction of scenes from multiple views with a projective camera, and the recognition of objects from a single or multiple views. The key reconstruction result is that projective reconstruction is possible from a set of more than one image of a point set. Methods of reconstruction using the fundamental matrix (from two views) and the trifocal tensor (from three views) are explored, as also is the possibility of self calibration of a camera, and consequent Euclidean reconstruction from three or more views. On the subject of object recognition from a single view, it is shown that by taking advantage of known, or hypothesized geometric structure of a scene being viewed, geometric invariants may be used for tasks of segmentation, grouping and object recognition.

Open PDF

Document Details

Document Type
Technical Report
Publication Date
Feb 28, 1997
Accession Number
ADA322738

Entities

People

  • Joseph L. Mundy
  • Richard I. Hartley

Tags

Communities of Interest

  • Air Platforms
  • Ground and Sea Platforms

DTIC Thesaurus Topics

  • Accuracy
  • Artificial Intelligence
  • Change Detection
  • Computer Programming
  • Computer Vision
  • Coordinate Systems
  • Databases
  • Geometric Forms
  • Geometry
  • Identification
  • Image Processing
  • Lines (Geometry)
  • Object Recognition
  • Recognition
  • Simplex Method
  • Three Dimensional
  • Two Dimensional

Readers

  • Computer Vision.