Cannot see the cylinder semantic model

[joexsyd]

I used Livox MID360 and OS64 LiDAR to record point cloud and IMU data in a real forest environment. I then used FastLIO2 as the odometry algorithm and ran tmux_single_outdoor_robot.sh to build a semantic map. However, I did not see the pink semantic cylinder model appear in the RViz window. There were warnings in the running window, indicating that there were too few tree-related points in the point cloud. Even when I tested with the 64-line OS LiDAR, the semantic cylinder still did not appear (the topic has already been subscribed to in RViz). Could you advise if there are any parameters I need to adjust? Also, what should I pay attention to when running my own data? I would be grateful for any advice or suggestions you could offer.

[ankitVP77]

I assume you are using the penn_smallest model weights to run the RangeNet++ LiDAR segmentation or are you using your own trained model? If you are using the penn_smallest weights then the tree trunk segmentation will be poor because this model was not trained on data from a forest but rather in an urban environment (like a parking lot). I have uploaded two more pre-trained model weights in the same Google Drive folder which are trained on forest data. I have also added a README document there explaining more about these models. Please take a look at those to understand which model maybe more suitable to you.

Once you figure out the front-end segmentation, please take a look at the cylinder_plane_modeller_params.yaml file to adjust the cylinder modeling parameters. There is also one parameter that is currently unfortunately hardcoded in the cylinder_plane_modeller.py file in the cluster function, currently on line 178, which you may also need to modify (we are still refactoring the code and this will be made a param later).

Finally, you mentioned that you are using FastLIO2. I just wanted to point out that we are using Faster-LIO and I am not sure how much backwards compatible our scripts are to FastLIO2. You may also need to check if the proper odometry topics are being supplied to the right nodes for the whole system to work.

I would suggest first debugging the semantic segmentation individually in RViZ to see if you are actually getting good segmented tree clouds. Start with debugging the /segmented_point_cloud_no_destagger topic published by the infer_node to see if trees are being segmented properly (they should have a different color to the rest of the point cloud) and then move to param tuning for the cylinder_plane_modeller.py script.