set_estimate on vertex results in seg fault

[lukasbuehler]

I experienced a similar issue as seen in: #44

However, the same issue hints at using
set_estimate_data instead of set_estimate, which works as expected.

If this is expected, the examples at https://github.com/miquelmassot/g2o-python-examples/blob/main
are wrong.

g2o-python Version 0.0.12

Example with Error

import g2o

p = g2o.VertexPointXY()
p.set_estimate([1., 2.]) # seg fault

Example without error:

import g2o

p = g2o.VertexPointXY()
p.set_estimate_data([1., 2.]) # works

[RainerKuemmerle]

At least from my end this is not expected.

As posted on more up to date wrapper code this works as expected.

I assume, we would need to check building wheel with the newer python binding.

[lukasbuehler]

I tried downgrading from Python 3.10 to Python 3.9, and also to g2o-python verision 0.0.11 and sadly always get the same results.

I tried it in a fresh Google Colab notebook, where it also crashes due to a segfault, I think the same thing is happening in #44 .

Notebook code:

!pip install g2o-python==0.0.12
import g2o
import numpy as np

v_point_xyz = g2o.VertexPointXYZ()
arr = np.array([1., 2., 3.])
v_point_xyz.set_estimate(arr)

Are there specific versions of wheel or other dependencies that are verified to work?

I don't think it is related to a wrong Numpy data type or wrong shape, because those throw normal errors.
And something else that is maybe interesting is that it is working for VertexSE3Expmap.

Code to reproduce:

v_se3 = g2o.VertexSE3Expmap()
pose_quaternion = g2o.SE3Quat(np.eye(3), np.zeros((3,), dtype=np.float64))
v_se3.set_estimate(pose_quaternion) # Works

Besides that, I have the same issue with edges and set_measurement, specifically:

edge = g2o.EdgeProjectXYZ2UV()
edge.set_vertex(0, ...) # ... are the existing vertices but omitted here for clarity
edge.set_vertex(1,...)

edge.set_measurement(keypoint) # seg fault

And using set_measurement_data on an Edge only takes in one Python float, which I don't know what it stands for, since in the C++ version in the documentation it takes in a double pointer and not a number, but since i cannot pass it the pointer from python i assume that this is another bug. Or can you help me understand what I can pass to set_measurement_data to get the code running?

Thank you for your help!

[RainerKuemmerle]

I assume that there are some issues in the older code of the g2o binding. I can try to fix the binding code under https://github.com/RainerKuemmerle/g2o/tree/pymem but do not see a direct way how to work on this repo.

As an intermediate step I looked into options to allow easier installation of the binding itself. See RainerKuemmerle/g2o#888

[lukasbuehler]

I was able to get it running by following the suggestions in #40 .
In summary I cloned the python-g2o repo (https://github.com/miquelmassot/g2o-python), then replaced the g2o subrepo there with your upstream repo (https://github.com/RainerKuemmerle/g2o/tree/pymem) and then I was able to install it locally using pip.

Thank you for your help and swift responses. It is even running on Python 3.12 now.

I'm in favor of not closing this issue since the problem still persists for the g2o-python PyPi package.