Photometric Bundle Adjustment for Vision-Based SLAM
C++ CMake
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
apps
cmake
config
data/kitti_init_poor
src
.gitignore
CMakeLists.txt
LICENSE
README

README

                             P H O T O  B U N D L E

This is the source code for
H. Alismail, B. Browning, and S. Lucey,
``Photometric Bundle Adjustment for Vision-Based SLAM,'' ACCV 2016

See additional information here as well:
https://www.cs.cmu.edu/~halismai/h_alismail_robotics_2016.pdf


                                  INSTALLATION

The code is tested on Linux (Debian, g++-4.9 and kernel 3.16.0-4-amd64)

Required packages:
- opencv
- ceres
- eigen3
- boost (program_options and filesystem)
- TBB [optional]


                                     USAGE

See the example code in apps/run_kitti.cc
The code is meant to be used as refinement on initial pose results obtained with
a visual odometry, or VSLAM system.

IMPORTANT: pose initialization is the *local* frame--frame pose. The code
returns the refined poses in the *global/cumulative* coordinate frame.


                                      TODO

The code is a research base and is very rough around the edges. It currently
does not run in real-time, but there are various possible improvements that
should improve the speed, which include:

- Pyramid implementation

- Refactoring the DescriptorError function to create a CostFunction per image
instead of per point. The current implementation (an objective per point) is
very inefficient due to thrashing the cache

- In addition to refactoring, implementing the CostFunction with an analytic
Jacobian should significantly improve performance. The current implementation
relies on ceres' automatic differentiation facilities which limit any
possibilities of exploiting SIMD instructions for a more efficient
implementation (the bottleneck is in patch interpolation)


There are also opportunities for improving the accuracy. These include:
- Proper selection of the characteristic scale of the Loss function
- Better pixel selection:
  - By selecting the smallest number of points sufficient to produce pose
  estimates robustly
  - By selecting the *best* points to constrain the solution


If you have questions/thoughts please contact me

Thu Dec  1 00:13:44 EST 2016
Hatem Alismail <hatem.alismail @ gmail>