Abstract:

Adjusting the bundles of lights projecting back from the 3D scene to the 2D images, known as bundle adjustment, creates accurate 3D locations in the structure from motion problem. This is a computationally expensive minimization step which is required to be recursively called if we want to reach accurate outputs from noisy sets of inputs and prevent the camera drift. This paper addresses a faster way to reconstruct large sets of images on a single computer by clustering the image data into smaller portions and reconstructing the whole scene in a hierarchical manner to reduce this step. Experimental results show that this hierarchical approach results in high quality 3D outputs, and further allows the use of parallel programming to speed up the overall process.

[PDF]