Image Processing in Catadioptric Planes: Spatiotemporal Derivatives and Optical Flow Computation

  • Authors:
  • Kostas Daniilidis;Ameesh Makadia;Thomas Bulow

  • Affiliations:
  • -;-;-

  • Venue:
  • OMNIVIS '02 Proceedings of the Third Workshop on Omnidirectional Vision
  • Year:
  • 2002

Quantified Score

Hi-index 0.01

Visualization

Abstract

Images produced by catadioptric sensors contain a significant amountof radial distortion and variation in inherent scale. Blindapplication of conventional shift-invariant operators or optical flowestimators yields erroneous results. One could argue that given acalibration of such a sensor we would always be able to removedistortions and apply any operator in a local perspectiveplane. In addition to the inefficiency of such an approach, interpolationeffects during warping have undesired results in filtering. In thispaper, we propose to use the sphere as the underlying domain of imageprocessing in central catadioptric systems. This does not mean that wewill warp the catadioptric image into a spherical image. Instead, wewill formulate all the operations on the sphere but use the samplesfrom the original catadioptric plane.As an example, we study theconvolution with the Gaussian and its derivatives and as well as thecomputation of optical flow in image sequences acquired with aparabolic catadioptric sensor.