Skip to content

Latest commit

 

History

History
90 lines (69 loc) · 4.12 KB

README.md

File metadata and controls

90 lines (69 loc) · 4.12 KB
Raw

Structure From Motion (SfM) reconstruction library

This library provides a toolset to perform SfM reconstructions on the Photo Tourism dataset (or datasets with similar formats). Performed as a "weekend" coding test.

Requeriments

This library is tested using Python 3.6.8. The full list of packages can be found in requirements.txt.

Dataset requeriments

This library is developed to be used in the Photo Tourism dataset or datasets with a similar format. The command line tool is given the path to the *.out file generated by Bundler and must follow the same file structure:

<dataset_folder>
├── images
├── list.txt
└── <bundle_file>.out

Basic usage

Some basic commands that can be performed using this library are:

  • Reconstruction using two images, view on a window and save the result on the file reconstruction.ply.
$  python sfm.py --path ~/Sources/coding_test/NotreDame/notredame.out --images 1 2 --colorize --view --save reconstruction.ply
  • Reconstruction using four images on a pairwise reconstruction, view on a window and save the result on the file reconstruction.ply.
$ python sfm.py --path ~/Sources/coding_test/NotreDame/notredame.out --images 1 2 3 4 --colorize --reconstruction_method pair --view --save reconstruction.ply
  • Reconstruction using four images on a bundle reconstruction, view on a window and save the result on the file reconstruction.ply.
$ python sfm.py --path ~/Sources/coding_test/NotreDame/notredame.out --images 1 2 3 4 --colorize --view --save reconstruction.ply
  • Reconstruction using four images on a bundle reconstruction, view on a window and save the result on the file reconstruction.ply, using an universal interface.
python sfm.py --path ~/Sources/coding_test/NotreDame/notredame.out --images 1 2 3 4 --demo --view --save reconstruction.ply

Parameters

The full list of parameters can be found using python sfm.py --help.

usage: sfm.py [-h] [--path PATH] [--images IMAGES [IMAGES ...]]
              [--reconstruction_type RECONSTRUCTION_TYPE]
              [--reconstruction_method RECONSTRUCTION_METHOD] [--view]
              [--colorize] [--save SAVE] [--demo]

Structure From Motion coding test uding the Photo Tourism dataset.

optional arguments:
  -h, --help            show this help message and exit
  --path PATH           Path to the *.out file containing the bundler data
  --images IMAGES [IMAGES ...]
                        Input images to the SfM algorithm (int indexes). The
                        method will fail if a number larger than the dataset
                        size is specified
  --reconstruction_type RECONSTRUCTION_TYPE
                        Type of reconstruction: 'dense' or 'sparse'. Defaults
                        to 'sparse'
  --reconstruction_method RECONSTRUCTION_METHOD
                        Method to perform the reconstruction: either 'pair' or
                        'bundle'. Defaults to 'bundle'
  --view                View the resulting point cloud in an X window
  --colorize            Colorize the point cloud
  --save SAVE           Save the resulting point cloud on the specified file
  --demo                Execute demo registration with proposed interface

Proposed interface

The global interface that can be used with any data source is the follwoing:

points, colors = sfm.sparse_reconstruction.sparse_bundle_reconstruction_data(images, pair_dict, camera_matrices, ks, colorize=True)

This interface needs the following parameters:

  • images List of numpy arrays with all the images.
  • pair_dict List of dictionaries with all the matches. For an example of how to create such list of dictionaries, see sfm.py.
  • camera_matrices List containing all the 3x4 camera matrices (intrinsics+extrinsics).
  • ks List containing all the intrinsic parameters for each image in the format [f, k1, k2].
  • colorize Optional parameter to indicate if the colors are also extracted. Defaults to True.

Changelog

V0.1 - Initial version with only sparse reconstruction working.