GTSfM: A distributed large-scale SfM library

John Lambert, Ayush Baid, Akshay Krishnan, Travis Driver, Ren Liu, Adi Singh, Frank Dellaert

A reconstruction of cameras and 3D points using 12 images from the Lund Door dataset

GTSfM is a global Structure from Motion (SfM) library which estimates 3D structure and camera motion from a collection of images. It is designed to support parallelized computation on a cluster, but can also be used on standalone devices.

Some of its key features are the following:

  • State-of-the-art deep feature extraction and feature matching for two-view relative pose estimation.
  • Robust rotation averaging, translation averaging and bundle adjustment using factor graphs, implemented with GTSAM.
  • A deep multi-view stereo pipeline to estimate dense 3D structure.
  • Support for parallelizing computation on distributed infrastructure using Dask.

The library is open-sourced, and can be used from the GitHub repository