ImportError for VO and Full SLAM Module

[donkeymouse]

Hello guys!
First of all, Thank you for such a wonderful work! As a fellow academic who researches in optical flow, this is a great asset to the community!

I have been trying to execute the VO and the Full-SLAM module following your guidelines you suggested.
I have installed all dependencies you mentioned (libopencv-dev, DBoW3, Ceres2 (all tests passed!), PyOpenGL) and think I managed to create the 'vo.so' and 'slam_full.so' files (with bunch of warnings displayed in the terminal window?).

I tried to run the VO demo first then execute the Full Slam demo, however, from both approaches I receive the following import errors.

"VO"

"ImportError: /home/hri/PycharmProjects/VOLDOR/VOLDOR/demo/pyvoldor_vo.cpython-36m-x86_64-linux-gnu.so: undefined symbol: _ZN2cv6String10deallocateEv"

"Full SLAM"

"ImportError: /home/hri/PycharmProjects/VOLDOR/VOLDOR/demo/pyvoldor_full.cpython-36m-x86_64-linux-gnu.so: undefined symbol: cusolverDnSetStream"

I am not sure of how to handle the following errors?

The environment I am currently working with is:

• Ubuntu 18.04
• OpenCv 3.2.0
• CUDA 11.3
  Any further information you need, I will be happy to provide

Again, thanks for your support

[htkseason]

Thanks for your kind words. The problem with your VO compling seems related to opencv. Did you install it from apt? Can you try using opencv 3.4.X follow the README?

[donkeymouse]

Hello Min! Thanks for your reply.

By following your instructions, I reinstalled OpenCV from source and was able to run the VO part.
(It seemed like I needed to reinstall my OpenCV after my recent upgrade of my GPU)

However, I am still unable to run the full-SLAM architecture. I am able to execute the setuplinux_full.py as you have instructed and can generate the .so and linux-gpu kernel files.

I still suffer from the following error: "ImportError: /home/hri/PycharmProjects/VOLDOR/VOLDOR/demo/pyvoldor_full.cpython-36m-x86_64-linux-gnu.so: undefined symbol: cusolverDnSetStream"

From looking at other issue posts, it seemed like there were frequent sources of errors due to incorrect installation of Ceres2 with SuiteSparse so I reinstalled the entire library again and was able to pass all the tests ( and have executed sudo make install)

I followed the instruction and install SuiteSparse through "sudo apt-get install libsuitesparse-dev"

After all this, I was stilll unable to execute the full_slam pipeline.

Do you have any idea on how to solve this?

Many Thanks

donkeymouse

[htkseason]

Can you run VO-only now?

[donkeymouse]

Hello Min. Yes, I can run VO-only now. However, I am still not able to run the Full-SLAM module part...

[htkseason]

Your current issue seems similar to this guy.
#17 (comment)
Could you please try add -l cusolver -l cublas to slam_py/install/setup_linux_full.py line 16.
If it does not work, maybe add cusolver, cusolver to the list in line 29.

It is strange since what is additionally needed in full-slam pipeline is only ceres, but this error points to cuda. Have you installed cuda toolkit?

[donkeymouse]

Hello Min, Thank you for your reply.

I have attempted both of your suggested methods (adding -l cusolver to line 16 and adding it to the list in 29) but the problem persists.

I have tried the following line for the first approach:

gpu_kernel_build_cmd = f'/usr/local/cuda/bin/nvcc --compiler-options "-shared -fPIC" {gpu_sources_cpp} {gpu_sources_cu} -l cusolver -l cublas -lib -o libgpu-kernels.so -O3 {nvcc_machine_code}'

I have cuda toolkit 11.4 installed into my machine.

Here is my output from running " nvcc -V"

Happy to provide more info if you need one.

Thanks alot

donkeymouse

[bytezh]

Hello Min, Thank you for your reply.

I have attempted both of your suggested methods (adding -l cusolver to line 16 and adding it to the list in 29) but the problem persists.

I have tried the following line for the first approach:

gpu_kernel_build_cmd = f'/usr/local/cuda/bin/nvcc --compiler-options "-shared -fPIC" {gpu_sources_cpp} {gpu_sources_cu} -l cusolver -l cublas -lib -o libgpu-kernels.so -O3 {nvcc_machine_code}'

I have cuda toolkit 11.4 installed into my machine.

Here is my output from running " nvcc -V"

Happy to provide more info if you need one.

Thanks alot

donkeymouse

Hello, I got the same error with you.

I found that the current ceres official website is ceres 2.1. I downloaded ceres 2.0 on github and solved this error