Open AccessMODERN METHODS OF BUNDLE ADJUSTMENT ON THE GPU
Author(s) -
R. Hänsch,
I. Drude,
O. Hellwich
Publication year - 2016
Publication title -
isprs annals of the photogrammetry, remote sensing and spatial information sciences
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.356
H-Index - 38
eISSN - 2194-9042
pISSN - 2196-6346
DOI - 10.5194/isprsannals-iii-3-43-2016
Subject(s) - intersection (aeronautics) , conjugate gradient method , gradient descent , acceleration , computer science , levenberg–marquardt algorithm , bundle adjustment , nonlinear system , bundle , descent (aeronautics) , nonlinear conjugate gradient method , algorithm , position (finance) , descent direction , mathematical optimization , mathematics , artificial intelligence , artificial neural network , physics , materials science , finance , classical mechanics , quantum mechanics , composite material , engineering , economics , image (mathematics) , aerospace engineering
The task to compute 3D reconstructions from large amounts of data has become an active field of research within the last years. Based on an initial estimate provided by structure from motion, bundle adjustment seeks to find a solution that is optimal for all cameras and 3D points. The corresponding nonlinear optimization problem is usually solved by the Levenberg-Marquardt algorithm combined with conjugate gradient descent. While many adaptations and extensions to the classical bundle adjustment approach have been proposed, only few works consider the acceleration potentials of GPU systems. This paper elaborates the possibilities of time and space savings when fitting the implementation strategy to the terms and requirements of realizing a bundler on heterogeneous CPUGPU systems. Instead of focusing on the standard approach of Levenberg-Marquardt optimization alone, nonlinear conjugate gradient descent and alternating resection-intersection are studied as two alternatives. The experiments show that in particular alternating resection-intersection reaches low error rates very fast, but converges to larger error rates than Levenberg-Marquardt. PBA, as one of the current state-of-the-art bundlers, converges slower in 50 % of the test cases and needs 1.5-2 times more memory than the Levenberg- Marquardt implementation.