search

PRBonn / lidar-bonnetal

Semantic and Instance Segmentation of LiDAR point clouds for autonomous driving
http://semantic-kitti.org
MIT License
959 stars 206 forks source link

Inference pipeline!! #38

Closed abhigoku10 closed 4 years ago

abhigoku10 commented 4 years ago

@Chen-Xieyuanli @niosus @stachnis @ovysotska @tano297 thanks for opensourcing the work i have few queries

1.i want to detect only stationary objects in the driving environment , can we use your pretrained model to give only those output , if so which secti0n of the code i have to modify 2.after the segmentation of the object class , how to obtain the cuboid bounding box of those object so that i can extract the area , distance and other params

  1. Can we test the model on lower lidar points/ring ? have you done any testing if so what is the accuracy drop 4.when we give custom data as input to the infer.py should it be also in sequences or it can be independent of it Thanks for the response !!!!
abhigoku10 commented 4 years ago

@Chen-Xieyuanli @tano297 @niosus any updates on these queries??

jbehley commented 4 years ago
  1. the pre-trained models do not distinguish between moving and non-moving objects.
  2. we provide here models for semantic segmentation, i.e., you will need to perform some clustering afterwards to get all points belonging to a single instance and use this to generate a bounding box.
  3. in our experience does this not work. There will be a significant drop in accuracy.
  4. the dataloader expects a certain folder structure. if you want to use a different structure then you have to replace the data loader or rewrite it.

We provide the code here for demonstration purposes and to give a starting point for your own development. We do not have the intention to cover all uses cases.

abhigoku10 commented 4 years ago

@jbehley thanks for the response . as you mentioned i shall make a pull request as i want to keep adding features for this . Thanks for open sourcing the work i was looking where to start from and urs gave me a starting point .

jbehley commented 4 years ago

there is no need to have a pull request per se. Only if you think that others could benefit from your effort.

Could I answer your questions and can we close this issue? (you can also re-open the issue if you feel something needs to be clarified.)