Image montage for constructing photorealistic virtual world from different real scene images

An image montage technique creates a photorealistic output image without artifacts by synthesizing multiple input images. We propose an image montage method to create a photo-realistic virtual scene that seems as if it exists in a real world by synthesizing real scene images captured in different places. The images to synthesize are appropriately selected from a large database of different scene images. We suppose that images in the database were captured in the same place by moving a camera to translate its 3D position and rotate its viewing direction. Our method first obtains lots of similar image pairs such that two images in each pair have similar regions as if they are the same scene part that shifts its position in respective images by a camera motion. Then, from the pairs, appropriate successive images are selected and arranged in a 3D virtual space such that every image is similar to both adjacent images in their overlapping regions to fake a camera motion of a walk-through in the space, which results in constructing a photorealistic virtual world. The similarity in the overlapping region is evaluated using gist and color histogram. The selection of the successive images is done by solving an optimization problem using a graph of images to search for appropriate closed paths, each of which satisfies a restriction of a 3D space. A final virtual scene image is synthesized by blending the overlapping regions of adjacent images in a desired closed path.