Geometry-Aware Image Completion via Multiple Examples

Guihang Wang, Xuejin Chen, Siyu Hu
CAS Key Laboratory of Technology in Geo-spatial Information Processing and Application System
University of Science and Technology of China


Algorithm overview



Abstract: When browsing through photographs taken during a trip, it can be a distressing discovery to find many other bystanders captured within the frame. A visually compelling snapshot preserves the desired subject in the foreground, and eliminates irrelevant persons or objects. In this paper, we present a new image completion algorithm that retains the geometric consistency of the scene using Internet photo collections. Our key insight allows us to successfully fill the missing regions with content derived from related images with distinct features, or from visible parts of the input image with repetitive textures. The final composition for the missing regions is guided by a confidence map combining the two types of example images. We validate our algorithm on a variety of scenes which are challenging for state-of-the-art completion techniques. The completion results successfully preserve the geometry consistency for complex structures in a wide range of depths.

Paper: [PDF 32.0M].
PG2014 Poster: [PDF 3.1M].
Bibtex: [Bibtex.txt].
PG2014 Short Papers: [Link].