MegBA: A High-Performance and Distributed Library for Large-Scale Bundle Adjustment

This repo contains an official implementation of MegBA.

MegBA is a fast and distributed library for large-scale Bundle Adjustment (BA). MegBA has a novel end-to-end vectorised BA algorithm which can fully exploit the massive parallel cores on GPUs, thus speeding up the entire BA computation. It also has a novel distributed BA algorithm that can automatically partition BA problems, and solve BA sub-problems using distributed GPUs. The GPUs synchronise intermediate solving state using network-efficient collective communication, and the synchronisation is designed to minimise communication cost. MegBA has a memory-efficient GPU runtime and it exposes g2o-compatible APIs. Experiments show that MegBA can out-perform state-of-the-art BA libraries (i.e., Ceres and DeepLM) by ~50x and ~5x respectively, in public large-scale BA benchmarks.

Version

• 2021/12/06 Beta version released! It corresponds to this paper.
• General version code release (Expected Dec 31 2021)
• memory-efficient version with implicit Hessian (TBD)
• analytical differential module, IMU factor, prior factor (TBD)

Paper:

• First Draft: https://arxiv.org/abs/2112.01349 (an updated version will be released by Dec 10)

Quickstart

Dependencies:

• C++14
• CMake (>= 3.15)
• CUDA Toolkit (with thrust) https://developer.nvidia.com/cuda-downloads
• Eigen (MegBA Beta depends on a modified version of Eigen under third_party/. Eigen-free version will be provided in the general release.)
• NCCL (if you need Distributed features) https://developer.nvidia.com/nccl/nccl-download

You can also easily install all dependencies with script: script

Demo with BAL dataset:

• Download any pre.txt.bz2 file from BAL Dataset: https://grail.cs.washington.edu/projects/bal/ and uncompressed.

• Compile

  mkdir build
  cd build
  cmake ..
  make -j4 BAL_Double

• Run the demo (Venice-1778)

  cd examples
  ./BAL_Double --name=Venice --world_size=2 --iter=100 --solver_tol=1e-1 --solver_refuse_ratio=1 --solver_max_iter=100 --tau=1e4 --epsilon1=1 --epsilon2=1e-10

  • world_size: number of GPUs available
  • iter: the maximal number of LM iteration
  • epsilon: threshold in LM
  • solver_tol: tolerance of solver (distributed PCG solver)
  • solver_refuse_ratio: early stop for the solver
  • solver_max_iter: the maximal iteration of solver
  • tau: the initial region

Notes for the practitioners

• Currently, MegBA implements automatic differentation only for generalizability. Please consider implementing your own analytical differentiation module.
• If you use devices without modern inter-device communication (i.e., NVLinks..), you might find the data transfer is the bottleneck.
• Empirically, we found it is necessary to customize the LM trust-region strategies and tune its hyper-parameters to further boost the performance.

Documentation

Under doc/ (Coming soon...)

Collaborate with Us

Please check here for MegBA's future plan.

If you are intereted in MegBA and want to collaborate, you can:

• Apply for an Internship at Megvii Research 3D, please send your resume to [email protected], with your expected starting date. (subject: 3D组CUDA实习生-Name) Unfortunately, now we are only able to host interns with work authorization in China.
• As an external collaborator (coding), just fork this repo and send PRs. We will review your PR carefully (and merge it into MegBA).
• As an algorithm/novelty contributor, please send an email to [email protected].
• Any new feature request, you can send an email to [email protected] as well. Note that it is not guaranteed the requested feature will be added or added soon

Contact Information:

• Jie Ren [email protected]
• Wenteng Liang [email protected]
• Ran Yan [email protected]
• Shiwen Liu [email protected]
• Xiao Liu [email protected]

BibTeX Citation

If you find MegBA useful for your project, please consider citing:

@misc{2021megba,
  title={MegBA: A High-Performance and Distributed Library for Large-Scale Bundle Adjustment}, 
  author={Jie Ren and Wenteng Liang and Ran Yan and Luo Mai and Shiwen Liu and Xiao Liu},
  year={2021},
  eprint={2112.01349},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}

License

MegBA is licensed under the Apache License, Version 2.0.