Technical reports

Synthetic image generation for a multiple-view autostereo display

Oliver M. Castle

October 1995, 184 pages

This technical report is based on a dissertation submitted April 1995 by the author for the degree of Doctor of Philosophy to the University of Cambridge, Wolfson College.

Abstract

Multiple-view autostereo displays provide true 3D images which offer depth cues (such as binocular stereo and motion parallax) that are not available with conventional monoscopic displays. This thesis addresses some of the problems of synthetic image generation for a multi-view autostereo display, with particular attention paid to the question of how rendering costs may be reduced by taking advantage of the similarities between the views.

A functional description of the prototype multi-view autostereo display device developed separately at the University of Cambridge sets the technological beckground of this research. The problems faced by synthetic image generation in general are reviewed next, of which visible surface determination is identified as the most important for multi-view stereo. A viewing model for multi-view stereo is then derived, building on experience with existing monoscopic and two-view stereoscopic viewing models.

Using this multi-view autostereo viewing model as a framework, two distinct approaches to multi-view stereo image synthesis are investigated. The first is an extension of conventional Z-buffer rendering methods, adapted to take advantage of the coherence between the views to eliminate redundant processing and share computational overheads whenever possible. The second, based on approximate stereo reprojection techniques, shares visible surface information between the views in an attempt to eliminate processing those parts of the scene considered unlikely to be visible in the final image, thus trading off image quality against rendering speed.

An experimental evaluation of these two techniques demonstrates that both are capable of producing multi-view stereo images at a lower cost per view than similar single-view rendering methods. The results indicate that the performance improvements of both algorithms increase with the number of views in the image, reaching asymptotic levels as the shared processing costs become relatively less significant compared with the overall rendering time. Of the two, however, the approximate algorithm appears to offer the better potential speedup, owing to the way in which it enables the effective depth complexity of the scene to be reduced.

Full text

PDF (171.6 MB)

BibTeX record

@TechReport{UCAM-CL-TR-382,
  author =	 {Castle, Oliver M.},
  title = 	 {{Synthetic image generation for a multiple-view autostereo
         	   display}},
  year = 	 1995,
  month = 	 oct,
  url = 	 {https://www.cl.cam.ac.uk/techreports/UCAM-CL-TR-382.pdf},
  institution =  {University of Cambridge, Computer Laboratory},
  doi = 	 {10.48456/tr-382},
  number = 	 {UCAM-CL-TR-382}
}