obtained, the calculated results of this current iterative step will be abandoned and
then a diagonal values will be modified automatically. Therefore, additional linear
equation computation is required to find the correct descent direction and step size.
Moreover, it is extremely like to drop into a larger local minimum using the LM solu-
tion, resulting in poor estimation. Furthermore, LM also has a linear convergence rate
problem with small residuals because it only uses the L-order information.
In the mathematical community, it has been demonstrated that the quasi-Newton
methods are effective to solve non-linear least square problems. Similar to the struc-
ture of LM, we recall several structured quasi-Newton methods (Zhou and Chen 2010)
with high convergence rates. The Broyden-Fletcher-Goldfarb-Shanno (BFGS) algo-
rithm is one type of quasi-Newton methods which is used in number computing (Li
and Fukushima 2001) and convolutional network model (Le et al. 2011), and yields
satisfactory results. In addtion, BFGS and L-BFGS (Liu and Jorge 1989) are also
used as solvers to solve optimization problems in the open source softwere – Ceres
(Agarwal, Mierle, and Others 2010). But, this software used those methods to obtain
the Hessian matrix directly instead of considering the special structure of the Hes-
sian metrix, imposing more computational burdens on machines. (Kaj Madsen 2004)
proposed a hybrid method based on both LM method and a Quasi-Newton one, in
which descent directions and descent steps were estimated by BFGS and a trust region
approach during Quasi-Newton steps. But, the experiments have demonstrated that
the speed of gradients decrease is slow. What’s more, less attention has been given to
the combination between the special structure and the BFGS method in solving BA
problems of large-dimension optimization by researchers in the photogrammetry and
computer communities.
This letter provides a sparse BFGS (sBFGS) method to solve BA problems. Similar
to GN and LM, the proposed method also uses a special structure to approximate the
Hessian matrix. Unlike the GN algorithm (higher-order terms of the Hessian matrix
are ignored) and LM algorithm (a simple damping item is added in the diagonal), the
sBFGS method replaces the damping item of LM of a sparse gain matrix computed
by the BFGS algorithm. This method is more accurate to approximate the Hessian
with a sparsity character, which retains information from higher-order terms by adding
damping values into the diagonal and off-diagonal; thus, an accurate descent direction
and a proper step can be obtained for iterative optimization. The results of the exper-
iment verify that the method can obtain a superlinear convergence rate for different
types of initial values and can drop into a smaller local minimum.
2. Background
2.1. BA problem
The goal of the BA problem in the photogrammetry field is to find optimal unknown
parameters by minimizing the re-projection residuals when image points are measured
for observation. Therefore, the cost function of BA problem can be expressed by the
mean square residual of the re-projection image points. Here, unknown parameters
consist of the vectors of camera poses (c) and 3-D points (p), in which each pose is
often represented by three Euler angles and three translation vectors, respectively. The
observation function, with which 3-D points are projected into image points based on
2