ozansener / dynfu

DynamicFusion is a dense Simultaneous Localization and Mapping (SLAM) system capable of reconstructing non-rigid deforming scenes.

Dynfu

Dynfu overview and report

Dependencies:

• CUDA-enabled CPU (Kepler or newer) with CUDA 7.5 or higher, required by Opt
• OpenCV 2.4.9 with new Viz module (only opencv_core, opencv_highgui, opencv_imgproc, opencv_viz modules required); make sure that WITH_VTK flag is enabled in CMake during OpenCV configuration

Implicit dependency (needed by opencv_viz):

• VTK 5.8.0 or higher (apt-get install on linux, for Windows please download and compile from www.vtk.org)

Future additions

• Surface fusion using PSDF for non-rigid surfaces
• Exending the warpfield by adding in new deformation nodes
• Regularisation for the energy function

References

The CPU warpfield solver is based on Ceres.

@misc{ceres-solver,
  author = "Sameer Agarwal and Keir Mierle and Others",
  title = "Ceres Solver",
  howpublished = "\url{http://ceres-solver.org}",
}

The GPU warpfield solver is based on Opt.

@article{devito2016opt,
  title={Opt: A Domain Specific Language for Non-linear Least Squares Optimization in Graphics and Imaging},
  author={DeVito, Zachary and Mara, Michael and Zoll{\"o}fer, Michael and Bernstein, Gilbert and Theobalt, Christian and Hanrahan, Pat and Fisher, Matthew and Nie{\ss}ner, Matthias},
  journal={arXiv preprint arXiv:1604.06525},
  year={2016}
}

The KD-tree is created using nanoflann.

@misc{blanco2014nanoflann,
  title        = {nanoflann: a {C}++ header-only fork of {FLANN}, a library for Nearest Neighbor ({NN}) wih KD-trees},
  author       = {Blanco, Jose Luis and Rai, Pranjal Kumar},
  howpublished = {\url{https://github.com/jlblancoc/nanoflann}},
  year         = {2014}
}