Wide-Baseline Foreground Object Interpolation Using Silhouette Shape Prior.

Area of Science:

• Computer Vision
• Computer Graphics
• Geometric Modeling

Background:

• Synthesizing intermediate views from widely spaced cameras (wide-baseline) is difficult.
• Foreshortening and occlusions create significant differences between reference images, complicating pixel correspondence.

Purpose of the Study:

• To develop a method for disambiguating the ill-posed problem of intermediate view synthesis in wide-baseline stereo.
• To ensure interpolated views are consistent with plausible object silhouette transformations.

Main Methods:

• Utilizing an object-specific prior, a nonlinear shape manifold learned from previous observations.
• Estimating the evolution of epipolar silhouette segments between reference views.
• Reconstructing epipolar plane images (EPIs) and aggregating them to form an EPI volume.

Main Results:

• The method successfully disambiguates intermediate view synthesis for wide-baseline configurations.
• Object topology is preserved in synthesized intermediate views.
• Effective handling of self-occluded regions and severe foreshortening effects.

Conclusions:

• The proposed approach enables robust intermediate view synthesis even with significant camera separation.
• The use of learned shape priors is crucial for managing challenges like foreshortening and occlusion.
• The method reconstructs EPIs and their continuum, facilitating a comprehensive view synthesis.