Context-driven composite stereo reconstruction

  • Authors:

  • M. Nguyen;R. Gong;Y.-H. Chan;P. Delmas;G. Gimel'farb

  • Affiliations:

  • The University of Auckland, Auckland, New Zealand;The University of Auckland, Auckland, New Zealand;The University of Auckland, Auckland, New Zealand;The University of Auckland, Auckland, New Zealand;The University of Auckland, Auckland, New Zealand

  • Venue:
  • Proceedings of the 27th Conference on Image and Vision Computing New Zealand
  • Year:
  • 2012

Quantified Score

Hi-index 0.00

Visualization

Abstract

With respect to the menagerie of possible observed 3D scenes, no algorithm today for reconstructing such a scene from a stereo pair of images is uniformly better than all the others by their accuracy and processing speed. Generally, appropriate stereo reconstruction algorithms should be selected in accord with a type, or context of the scene and in many cases different parts of the same scene could be reconstructed most accurately by using different algorithms. To qualitatively explore this problem, we collected a database of more than 1,500 stereo pairs of natural and artificial indoor and outdoor 3D scenes, arranged into 25 types, such as animals, bars, city roads, city trees, classrooms, coasts, corridors, etc. The images were processed with two algorithms: the 2D graph-cut stereo (2DGCS) and the 1D belief propagation stereo (1DBPS) -- with automatically estimated parameters. The obtained depth maps were visually evaluated and compared by a number of independent human observers. Although in literature the 2DGCS is usually considered as more accurate than the 1DBPS, its depth maps were preferred by the observers in these experiments only for about 58% of the images and 15 out of the 25 types of scenes. The fast and easily parallelised 1DBPS restores smooth continuous curved surfaces but with noisy object boundaries due to horizontal streaks, whereas the much slower and intrinsically sequential 2DGCS returns flattened depth maps with distinctive object boundaries. Based on these results, we implemented context recognition to examine an input scene and allocate the most suitable algorithm and proposed to combine the 2DGCS and 1DBPS into a composite stereo reconstruction technique, which is qualitatively superior than each individual algorithm and is able to return better results and with good speed.