We provide both C++ and Matlab code that estimates where a picture was taken, giving a single pair of 3D local geometry structure (differential geometry) and its projection in a 2D image. Two particular cases are common
- A pair of 3D-2D of SIFT features (the 3D is just the SIFT center with reconstructed 3D SIFT orientation from two other images)
- A pair of 3D-2D point correspondences, where the points belong to curves (e.g., 2D-3D edgels, corners, or junctions)
The algorithm, P2PT, supersedes the usual P3P by using only two features (each with a direction). This also solves a local pose problem which can be integrated along curves to provide a more global estimate or a matching algorithm. The run time of the C++ code is in the same order as the best P3P implementation available (tens of microseconds).
A journal version is available at: https://ieeexplore.ieee.org/document/9057738
Camera Pose Estimation Using Curve Differential Geometry, IEEE Transactions on Pattern Analysis and Machine Intelligence - PAMI, 2020, Ricardo Fabbri, Peter J. Giblin and Benjamin Kimia.
This is an improvement over the research code originally written for the paper:
R. Fabbri, P. J. Giblin, B. B. Kimia, "Camera Pose Estimation Using Curve Differential Geometry", ECCV 2012, Firenze, Italy (Lecture Notes in Computer Science)
PDF and bibtex available at: http://multiview-3d-drawing.sourceforge.net
This work was developed at Brown University, University of Liverpool and Rio de Janeiro State University.
Main Function
Matlab
matlab/rf_pose_from_point_tangents_root_find_function_any.m
This function will return all possible (Rotation, Translation) solutions for a given pair of 3D-2D point-tangents (oriented points). These can then be tested within standard RANSAC by you to keep only the one with the most inliers.
A simple demo with random data can be found in
matlab/demo.m
C++
See cmd/diffgeom2pose.cxx
Disclaimer
C++ version
This code is meant for production and is currently being incorporated to mainstream SfM pipelines such as OpenMVG and Colmap.
Matlab version
The matlab code is reseach code writen in a "lab" language (Matlab) and, despite
extensive experiments from our PAMI'20 paper showing it is robust and reliable,
one shouldn't expect it to be ready for production. A hardened version of this
code is the C++ implementation available in the folder p2pt/. The implementation
produces the results in the paper, and may well be very useful, but it may still
be hugely improved (as we have been doing for the C++ version). Oh, and don't
forget: please cite this paper ;^)
Contact
Please contact Ricardo Fabbri rfabbri@iprj.uerj.br or any of the other authors for requests.
This work was developed at Brown University, University of Liverpool and State University of Rio de Janeiro.
