brian-h-wang / ldls Goto Github PK

@brian-h-wang hi i have few queries regarding the few hyperparams used
Q1. for kitti dataset you have used calibrations matrix for projection of points , if for custom data i have no calibration file is there any way i can modify the modules
Q2 To connect 3D points you use K-means which for kitti is 10 how did you come up with this param , did you use same value for both velodyne-64 and velodyne-16
Q3 To perform label diffusion you use either convergence or max iterations how to come up with this value
Q4. Besides semi supervised graphnn did u try some other methods like dbscan, ecludiean clustering if so can you share that
Q5 you had cited "“Fusion of images and point clouds for the semantic segmentation of large-scale 3D scenes based on deep learning,”" did you the refined segmentation mentioned

THanks a lot in advance

@brian-h-wang hi thanks for sharing this wonderful code , i have few queries

1. can we perform the inference on CPU , is there CPU code available
   2.any instance segmenation or sematic segmentation model output can be used ?
   3.what is the performance and accuracy of the of the work

Thanks in advance

@brian-h-wang thanks for open sourcing your work which is saver for me currently !! i have few queries
Q1. Is semi supervised graph based method dependent on the density of points in the point cloud
Q2. Which part of the code performs graph creation and label diffusion can you please point it out
Q3.Can we use LDLS for static objects like trees , buildings ,poles and lanes
Q4 what is inference time for the whole code follow
Q5.Which part of the code performs the proper segmentation by outlier removal
Q6. Compared to Kitti dataset point density my custom dataset has less points so can i use LDLS

Thanks for the response

Hi
Thank you for this repo

This is not exactly an issue from your end. I just wanted to know if you guys would know the source of this error -
double free or corruption (!prev)

I am trying to run your repo on waymo's dataset and made minor changes to make it run.
https://www.diffchecker.com/zOxkARuM

Using TF2.2 and Ubuntu 18.04.

This result is amazing, how can I apply it to my special scene? I currently have an intel L515 radar sensor.

@brian-h-wang hi i have few queries on the inference time
Q1. in you notebook we get timing "363 ms ± 21.9 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)" what does this statement infer
Q2. I have obtain LDLS in real time inference which are of optimization i need to conc on since on 2080 Ti on a given frame which is sparse point cloud i am timing of "160ms"
Q3 By reducing the distance i.e along x axes which would have less point cloud will be able to
obtain high inference time
Q4 in Segmentation.py in line 385 you have mentioned "TODO: Check if omitting this is faster later." what is supposed to be done in here
Q5. can we use create_graph() for different class objects which might reduce the timing ??
Thanks in advance