Capture and Dewarping of Page Spreads with a Handheld Compact 3D Camera

Abstract

This paper describes a system for capturing images
of a book with a 3D stereo camera which performs
dewarping to produce output images that are flattened. A
Fujifilm consumer grade 3D camera (FinePix W3) provides
a highly mobile and low cost 3D capture device. Applying
standard computer vision algorithms, the camera is calibrated
and the captured images are stereo rectified. Due to technical
limitations, the resulting point cloud has defects such as
splotches and noise, which make it hard to recover the precise
3D locations of the points on the book pages. We address this
problem by computing curve profiles of the depth map and
using them to build a cylinder model of the pages. We then
generate a mesh M1 on the source image and project this into
a mesh M2 on the cylinder model in virtual space. Finally, the
mesh M2 is flattened and the pixels in M1 are interpolated and
rendered via M2 onto the output image. We have implemented
a prototype of the system and report on some preliminary
evaluation results.