OpenMVG offers a full set of functionalities towards Multiple View Geometry Reconstruction. Part of its library includes a complete Structure from Motion pipeline.
To estimate the camera poses (without structures i.e. no sparse or dense pointcloud), we can shorten the pipeline provided tutorial_demo.py to only run the following processes:
openMVG_main_SfMInit_ImageListing
openMVG_main_ComputeFeatures
openMVG_main_ComputeMatches
openMVG_main_IncrementalSfM --OR-- openMVG_main_GlobalSfM
Completing this pipeline, we will obtain `sfm_data.bin` which could be named differently according to your running command. To produce the readable results, we will need to excecute `openMVG_main_ConvertSfM_DataFormat`. The command can be executed like below:
cd $OPEN_MVG_BUILD/Linux-x86_64-RELEASE
openMVG_main_ConvertSfM_DataFormat -i sfm_data.bin -o sfm_data.json -V -I -E
where sfm_data.bin is the SfM binary result produced by OpenMVG which will be converted to a json file named as sfm_data.json, hence -o, and -V -I -E are entered such that views, intrinsics and camera poses (extrinsics), in respective order, will be exported.
The views section contains the image file names and camera's centre prior (i.e. when GPS data are read from exif).
The intrinsics section contains the camera intrinsics i.e. focal length, priciple points, and distortion coefficients.
The extrinsics section contains the camera rotation matrix R and the camera center C, in which we can derive the perspective projection matrix as P = K[R|t] in which t=-RC and K is the intrinsics matrix.
In this repository, a utility to extract camera pose [R|t] for each image and the camera intrinsics are extracted and stored as csv and txt files for further processing.
The code is developed in C++ using MS Visual Studio 2012 BUT there is no reference to any MFC specific functions. It uses RapidJson to read the content from the json file. However, downloading RapidJson is not required as all the necessary files are already included in this project under the folder "external".
After building, it can be run on Windows by using command
openmvg_pose_extraction FULL_PATH_TO_SFM_DATA.json. For example,
openmvg_pose_extraction D:\\OpenMVG_Results\\sfm_data.json
And running on Linux by using command
openmvg_pose_extraction /home/usr/source/OpenMVG_Results/sfm_data.json
Notice the slash (that is \\ for Windows and / for Linux) as the difference between the two operating systems. This is because the program will have to execute the mkdir PATH\\RT or mkdir PATH/RT command to create a folder named RT under the path where the file sfm_data.json is located.
An RT folder is created under PATH where sfm_data.json is located. The folder contains text files named as "<Image_file_name>.txt" each of which stores a 3x4 matrix corresponding to [R|t]. It also produces 2 CSV files: intrinsics.csv and poses.csv. intrinsics.csv have the intrinsics parameters for each camera specified by intrinsics_id. poses.csv contains camera poses.