The result of five_point_ransac() is not fixed about recoverpose.jl HOT 8 CLOSED

Most likely due to different random samples being drawn from data each time you run it.
This causes the algorithm to converge to different results.

In practice, however, I find this algorithm to be quite noisy. If you're using it for pose recovery over several frames, I'd recommend using something like bundle adjustment on top of it to optimize poses once you've recovered them.

from recoverpose.jl.

I'm doing self-calibration for multiple cameras.
previous_points, current_points K are all pre-made data.

n_inliers, (E, P, inliers, best_repr_error) = five_point_ransac(previous_points, current_points, K, K, GEEV4x4Cache())
run several times, E,P is not same.

And i try the essential_ransac(previous_points_pd, current_points_pd, focal_sum; threshold = 1.0) have error:
ERROR: LoadError: BoundsError: attempt to access Tuple{SMatrix{3, 3, Float64, 9}, Vector{Bool}, Float64} at index [4]
Stacktrace:
[1] indexed_iterate(t::Tuple{SMatrix{3, 3, Float64, 9}, Vector{Bool}, Float64}, i::Int64, state::Int64)
@ Base ./tuple.jl:86
[2] camera_1()
@ Main ~/codes/AutoCab.jl/src/test_2.jl:69
[3] top-level scope
@ ~/codes/AutoCab.jl/src/test_2.jl:99

from recoverpose.jl.

Looks like the error is in your code and I'm assuming you're trying to unpack the returned result.
There's error in documentation for essential_ransac. It returns n_inliers, (E, inliers, error), without P projecton matrix,
but looks like you are trying to unpack it to n_inliers, (E, P, inliers, error).

from recoverpose.jl.

n_inliers, (E, inliers, error) = essential_ransac(previous_points_pd, current_points_pd, focus_sum; threshold = 1.0)
is ok !!!

but the doc is bad:

from recoverpose.jl.

n_inliers, (E, inliers, best_repr_error) = essential_ransac(previous_points, current_points, focal_sum; threshold = 1.0)
best_n_inliers, P_res, best_inliers, best_error = recover_pose(E, pixels1, pixels2, K1, K2; threshold = 1.0)

Is this better than five_point_ransac() ??

from recoverpose.jl.

Agree, the documentation can be improved.
I find the five_point_ransac to be actually better, because it evaluates both essential matrix and recovered poses from it, while essential_ransac only evaluates essential matrix, from which you then evaluate only 4 possible poses.

from recoverpose.jl.

I see, thanks for your reply !!

from recoverpose.jl.

The documentation is fixed, thanks for pointing out. Closing this, feel free to re-open though.

from recoverpose.jl.