Description
This is an implementation of a sparse Levenberg-Marquardt optimization
procedure and several bundle adjustment modules based on it. There are three
versions of bundle adjustment:
1) Pure metric adjustment. Camera poses have 6 dof and 3D points have 3 dof.
2) Common, but adjustable intrinsic and distortion parameters. This is useful,
if the set of images are taken with the same camera under constant zoom
settings.
3) Variable intrinsics and distortion parameters for each view. This addresses
the "community photo collection" setting, where each image is captured with
a different camera and/or with varying zoom setting.
There are two demo applications in the Apps directory, bundle_common and
bundle_varying, which correspond to item 2) and 3) above.
The input data file for both applications is a text file with the following
numerical values:
First, the number of 3D points, views and 2D measurements:
<M> <N> <K>
Then, the values of the intrinsic matrix
[ fx skew cx ]
K = [ 0 fy cy ]
[ 0 0 1 ],
and the distortion parameters according to the convention of the Bouget
toolbox:
<fx> <skew> <cx> <fy> <cy> <k1> <k2> <p1> <p2>
For the bundle_varying application this is given <N> times, one for each
camera/view.
Then the <M> 3D point positions are given:
<point-id> <X> <Y> <Z>
Note: the point-ids need not to be exactly from 0 to M-1, any (unique) ids
will do.
The camera poses are given subsequently:
<view-id> <12 entries of the RT matrix>
There is a lot of confusion how to specify the orientation of cameras. We use
projection matrix notation, i.e. P = K [R|T], and a 3D point X in world
coordinates is transformed into the camera coordinate system by XX=R*X+T.
Finally, the <K> 2d image measurements (given in pixels) are provided:
<view-id> <point-id> <x> <y> 1
See the example in the Dataset folder.
Performance
This software is able to perform successful loop closing for a video sequence
containing 1745 views, 37920 3D points and 627228 image measurements in about
16min on a 2.2 GHz Core 2. The footprint in memory was <700MB.
Requirements
Solving the augmented normal equation in the LM optimizer is done with LDL, a
Cholsky like decomposition method for sparse matrices (see
http://www.cise.ufl.edu/research/sparse/ldl). The appropriate column
reordering is done with COLAMD (see
http://www.cise.ufl.edu/research/sparse/colamd). Both packages are licensed
under the GNU LGPL.
This software was developed under Linux, but should compile equally well on
other operating systems.
-Christopher Zach (chzach@inf.ethz.ch)
News for SSBA 2.0
* Added a sparse LM implementation (struct ExtSparseLevenbergOptimizer)
handling several least-squares terms in the cost function. This is useful
when several types of measurements (e.g. image feature locations and GPS
positions) are available. See Apps/bundle_ext_LM.cpp for a simple demo.
* Changed the default update rule for the damping parameter lambda to a
simpler one (multiply and divide lambda by 10, depending on the cost
function improvement). This seems to work better that the more complicated
rule used before. * Fixed a trivial, but important bug in cost evaluation after the parameter
update.
/*
Copyright (c) 2011 Christopher Zach, Computer Vision and Geometry Group, ETH Zurich
This file is part of SSBA-2.0 (Simple Sparse Bundle Adjustment).
SSBA is free software: you can redistribute it and/or modify it under the
terms of the GNU Lesser General Public License as published by the Free
Software Foundation, either version 3 of the License, or (at your option) any
later version.
SSBA is distributed in the hope that it will be useful, but WITHOUT ANY
WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR
A PARTICULAR PURPOSE. See the GNU Lesser General Public License for more
details.
You should have received a copy of the GNU Lesser General Public License along
with SSBA. If not, see <http://www.gnu.org/licenses/>.
*/