The RGBD camera running effect under the TUM data set was amazing.But below I encountered a new problem, I hope you can help.

[dansonZhang]

I tried running PLVSII again, and the RGBD camera running effect under the TUM data set was amazing. I admire it very much.

But below I encountered a new problem, I hope you can help.

I found that in Stereo mode or Stereo-Inertial mode, only line features can be added (other new feature parameters are set to 0). When I add new feature parameters such as PointCloudMapping.on: 1, the system will crash with a segmentation fault. Is this caused by my problem? Some detailed documents below. Run command: ./run_tum_vi_stereo.sh or ./run_tum_vi_stereo_inertial.sh Terminal output: terminal_output_run_tum_vi_stereo.txt or terminal_output_run_tum_vi_stereo_inertial.txt YAML file: TUM-VI_rectified.yaml(only modified PointCloudMapping.on: 1)

I found that when the stereo mode of PLVSII only turns on the point and line features, in the dataset-outdoors8_512_16 sequence of the TUM-VI dataset, after leaving the building, the trajectory will drift circularly at a constant speed (as shown in the figure stereo_TUM -VI-dataset-outdoors8_512_16_point_line_1.png, stereo_TUM-VI-dataset-outdoors8_512_16_point_line_2.png). It seems to be caused by fisheye camera distortion.

Similar to the second doubt, I found that the stereo-inertial of PLVSII, when only the point and line features are turned on, in the dataset-outdoors8_512_16 sequence of the TUM-VI dataset, is in a specific place (as shown in the figure stereo-inertial_TUM -VI-dataset-outdoors8_512_16_point_line_1.png, stereo-inertial_TUM-VI-dataset-outdoors8_512_16_point_line_2.png) will cause the IMU to run away.

[dansonZhang]

terminal_output_run_tum_vi_stereo.txt terminal_output_run_tum_vi_stereo_inertial.txt

[dansonZhang]

stereo_TUM-VI-dataset-outdoors8_512_16_point_line_1.png:

stereo_TUM-VI-dataset-outdoors8_512_16_point_line_2.png:

stereo-inertial_TUM-VI-dataset-outdoors8_512_16_point_line_1.png:

stereo-inertial_TUM-VI-dataset-outdoors8_512_16_point_line_2.png:

[luigifreda]

Hi, thanks for the post.

Some preliminary questions:

• Can you please share how you installed PLVS? Did you use a local install of opencv as suggested?
• Does your computer have a GPU?
• Can you share the log you obtain by running git diff >& /tmp/log.txt?

1

I found that in Stereo mode or Stereo-Inertial mode, only line features can be added (other new feature parameters are set to 0). When I add new feature parameters such as PointCloudMapping.on: 1, the system will crash with a segmentation fault.

I don't see this problem both on Euroc and TUM-VI. I can enable PointCloudMapping.on: 1 without problems. Can you test on Euroc if you have the same problem? If yes, can you enter in Thirdparty/libelas-gpu/build and run ./test_gpu and copy-paste here what you get as stdout?

I see you are getting the crash problem on the TUM-VI slide3 dataset.

• Can I ask why you selected that specific dataset to test point cloud mapping that is based on stereo depth estimation?
• If there is a specific reason, what's your thought about static stereo matching applied to stereo pairs like this?

I checked the dataset and nothing got wrong.

2

I found that when the stereo mode of PLVSII only turns on the point and line features, in the dataset-outdoors8_512_16 sequence of the TUM-VI dataset, after leaving the building, the trajectory will drift circularly at a constant speed (as shown in the figure stereo_TUM -VI-dataset-outdoors8_512_16_point_line_1.png, stereo_TUM-VI-dataset-outdoors8_512_16_point_line_2.png). It seems to be caused by fisheye camera distortion.

I reproduced this problem. I solved it by using the setting System.thFarPoints: 20.0 that ignores points that are too far away (like "sky" points). I am pushing some small changes that allow to correctly load that setting.

That being said, I think a couple of notes are in order. In TUM-VI, using a rectified stereo pair allows to use the lines and volumetric mapping. However, this requires a pre-processing (that acts as a pre-filter on the images during the rectification) and reduces the FOV of the camera. In general, these 2 factors may affect both SLAM robustness and accuracy.