Pyramid-Based 3D Reprojection Computations

Summary

With the current limitations of desktop computers in terms of memory, disk storage and computational power, scientists face several challenges when inspecting terabyte-sized (TB) images. These challenges result from the following:

• TB images are typically stored as thousands to millions of smaller files that have to be rendered seamlessly and interactively;
• Images need to be inspected with their surroundings of spatial, temporal and spectral information, and the neighboring image information has to be viewed from multiple viewpoints;
• Images need to be shared and then accessed by a team of experts.

For volumetric images that are too large for viewing interactively in a browser, our objective is to use the Deep Zoom representation and pre-compute multiple view points of a 3D data set. Figure 1 illustrates the computations needed for converting image representations and for changing view points. We have benchmarked algorithms for such conversions on a Hadoop cluster and on a desktop. In addition, we designed a theoretical model for cluster node utilization in order to request optimal distributed computing resources for such computations based on the parameters of input data.

Description

When handling very large images (larger than gigapixel images), there is a need to build a multi-resolution pyramid representation for fast rendering. In addition, the pyramid representations are built for multiple orthogonal view points to expedite rendering. We address this problem by designing five re-projection algorithms for three-dimensional (3D) images to enable their interactive visualization from multiple orthogonal viewpoints using the Deep Zoom pyramid representation. TB-sized 3D images consisting of [x, y, z or t or λ] are represented as an ordered set of either 2D cross sections [x, y] or Deep Zoom pyramids and then re-projected to [x, z] or [y, z] sets.

Five re-projection algorithms for the two 3D image representations are analyzed in terms of their computational time and memory complexities on a single machine and on a computer cluster with the Hadoop platform. We introduce a computational node utilization coefficient that relates algorithmic memory complexity and parameters of computational nodes.

Our contributions lie in the following:

• designing a new approach to 3D re-projection from a set of Deep Zoom pyramids,
• characterizing computational memory complexities of re-projection algorithms, and
• maximizing the speed-up of 3D re-projection computations on Hadoop computer clusters.

Major Accomplishments

The paper: Peter Bajcsy, Antoine Vandecreme, Mary Brady, “Re-projection of Terabyte-Sized Images, “ IEEE International Conference on Big Data, October 6-9, 2013, Santa Clara, CA, USA (poster)(download pdf)