A question:about camera pose optimization!

[ccy-ustb]

Hello, I am very interested in your work and would like to make some applications based on line map, such as visual localization..
Now that I have the initial pose and some 2D-3D line correspondences, how can I construct optimization problems to optimize the initial pose? （In the case of using only line correspondences instead of point line union）
Looking forward to your reply and wishing you a happy life！

[MarkYu98]

There is a ceres-based optimizer LineLocEngine:

limap/limap/optimize/line_localization/lineloc.h

Line 22 in 8450057

class LineLocEngine {

that only takes line correspondences and optimizes the initial camera pose. Take a look at limap.optimize.line_localization.solve_lineloc() and solve_lineloc_merged() for examples of how to use the optimizer from Python. You could also just use the point-line joint optimizer (JointLocEngine) while passing empty lists for point correspondences.

On a higher level, the heart of all our visual localization experiments is the function limap.estimators.pl_estimate_absolute_pose():

limap/limap/estimators/absolute_pose/pl_estimate_absolute_pose.py

Line 7 in 8450057

def pl_estimate_absolute_pose(cfg, l3ds, l3d_ids, l2ds, p3ds, p2ds, camera, campose=None,

.
When ransac.method is set to None in the configs, the localization is done using the optimizer mentioned over the initial pose; otherwise (hybrid) RANSAC is employed to robustly estimate the camera pose (in which case the initial pose is irrelevant). As line correspondences can be noisy, a RANSAC estimation might be preferred. Again, the joint optimization/hybrid RANSAC can be used with only line correspondences by passing empty lists for point correspondences.

We are currently working on documentations to better explain these functionalities in detail :)