pbias / lunet Goto Github PK

Hi there,

I am a researcher who has been working with your code for a few weeks trying to get together a pipeline to feed ROS Lidar point clouds through your network in real-time, and then export the result back to a ROS pointcloud.

After spending quite abit of time putting together a working pipeline, I realised that due to the slow nature of your pointnetize function (about 0.5s processing time), which as far as I can tell is required to compute the neighbours to feed into your network, I dont think real time operation on data which is not pre-processed is possible as you tout in your paper.

Am I missing something here? Or am I just out of luck on this one?

Thanks for your time

@mbredif @pbias thanks for sharing your code , i have following few queries
Q1. can we use the inference on custom dataset ? is so what is the changes to be made
Q2. Does sematic segmentation performed on other object class like tree building poles and barrier classes
Q3. to train the custom dataset is it should be in semantic kitti format ?
Q4 can you share your pretrained model to test he model on custom dataset and transfer learning
Q5 what is the performance achieved on live data

THnaks in advance

Hi,
Is the example config given in this repo the same as used for the paper results? In the paper I could not find anything about the pointnetize part. However, in the config given here it is enabled. Could you give me some insight about this?

Thanks!

Hi
Thanks for the great work and sharing.

Right now i'm VLP16 own dataset.
can i use it in LU-net or do i need to change the code accordingly?.

because Squeezeseg mentioned as its accepts only Kitti dataset

hi, im new in DL, hope u can give me some advice.

I saw method that LiDAR points mapped to the coordinates x,y in your paper

I can't find out this fuction which is make 3D coordinate to 2D, or I can find in data_loader.py

Nice work!

Any possibility to release the pre-trained model which can reproduce the results in your paper?