Introduction

This site concerns posest, a C/C++ library for 3D pose estimation from point correspondences that is distributed as open source under the GNU General Public License (GPL). Pose estimation refers to the computation of position and orientation estimates that fully define the posture of a rigid object in space (6 DoF in total). The computation is based on a set of known 3D points and their corresponding 2D projections on an imaging sensor. The library estimates the relative motion between the 3D points and the camera. Depending on the application, it can estimate object pose when the camera is stationary and the points originate from a moving object, or camera pose when a camera moves freely in a mostly stationary scene. Single or binocular camera systems are supported. Image points typically originate from local features, however posest is oblivious to their origin. Mismatched 3D-2D point pairs (i.e. outliers) are tolerated via robust regression techniques. Pose is estimated by optimizing physically meaningful geometric criteria.


The development of posest has been partially supported by the EC FP7 programme under grant no. 270138 DARWIN.


Technical Overview

posest includes monocular and binocular pose estimation variants. Binocular pose estimation offers increased accuracy at the cost of increased computing time. The approach adopted in each case is outlined below; more details can be found in our related ICVS'13 publication which is available here [bibtex].

  • Monocular robust pose estimation
    1. Preliminary pose estimation: A P3P or P4P solver is embedded in a RANSAC framework that uses the redescending M-estimator sample consensus (MSAC) cost function for hypotheses scoring. RANSAC determines the best scoring hypothesis and classifies correspondences into inliers and outliers. In addition to pose, the P4P solver can also yield focal length estimates (i.e., it is actually a P4Pf solver).
    2. Non-linear refinement: The preliminary pose is refined by minimizing the cumulative reprojection error for all inliers. To mitigate the influence of mislocalized 2D points, a M-estimate of the reprojection error rather than the squared Euclidean norm is minimized. This minimization is carried out iteratively with the Levenberg-Marquardt algorithm initialized with the preliminary pose and using analytic Jacobians.
    3. Global optimization: When the non-linear refinement is initiated far from the true minimum, it runs the risk of getting trapped to a local minimum. To counter this, multi-start global optimization can (optionally) be employed to explore the pose space using multiple local optimizations each initialized with a different starting point.

  • Binocular robust pose estimation
    1. Extrinsic calibration: The two cameras are calibrated extrinsically with the aid of a calibration grid. This procedure is performed offline as the calibration is assumed to remain fixed.
    2. Monocular pose estimation: Pose estimation is performed individually for each of the two cameras, as described above.
    3. Joint pose refinement: A single pose is estimated for the binocular pair, enforcing the constraint that the relative pose of the employed cameras is known. The estimation employs the inliers of each monocular estimation and is based on the minimization of the binocular reprojection error in both images. Such an approach circumvents the error-prone reconstruction of points via triangulation and does not limit the baseline of the two views nor calls for sparse feature or 3D point matching.
More techical details are provided in the FAQ and section Available Functions. Included demo programs posest_demo.c and binocposest_demo.c provide working examples of using posest. A MEX-file interface for using posest from within matlab is also provided.
It is stressed that posest does not include any means for detecting and matching point features between the 3D world and images. Such functionality can be supplied by other software such as, for example, Lowe's SIFT detector & descriptor, or the VLFeat and OpenSURF libraries.


Available Functions

posest provides two primary user-callable functions for performing pose estimation. Function prototypes and a brief explanation of their arguments is provided below. Some general conventions followed by both functions are as follows. Points (2 or 3D) are provided as 2D arrays where rows correspond to individual vectors and columns to vector coordinates. For example, the two 3D vectors (10., 20., 30.) and (70., 80., 90.) are laid out in memory as

10. 20. 30.
70. 80. 90.
which is of course equivalent to
10. 20. 30. 70. 80. 90.
2D matrices such as intrinsics or camera matrices should be supplied using the standard (i.e., row-major) C convention. Both functions employ the same return codes.
The included posest_demo.c sample program gives a working example of pose estimation from user-supplied data.

posest()

.:: toggle display ::.

posestBinoc()

.:: toggle display ::.


Contact Address

If you find this package useful or have any comments/questions/suggestions, please contact us at

Be warned that although we try to reply to most messages, it might take long to do so.
In case that you use posest in your published work, please cite this paper and acknowledge this site.



hits since Thu Oct 9 12:04:27 EET 2014