Giters

loicland / superpoint_graph

Large-scale Point Cloud Semantic Segmentation with Superpoint Graphs

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Test on my custom_dataset

zcy1065670505 opened this issue · comments

commented

HI,
When I use SPG to test my custom dataset,Do I need to make any other changes except what you mentioned in readme.
In fact, I have tested the custom dataset (LIDAR highway scene) according to your instructions, but there is always one class missing in the test results (the lane line is missing). I do this.
First of all, my data does not have RGB(only x y z intensity). I replace the RGB channel with the intensity of lidar (copy three times),then I use the intensity as additional information besides geometric information for partition.
Secondly, my datasets are all labeled, and there are no unlabeled points,what other changes do I need to make to get the correct segmentation results.
Thank you!
the gt and pred are as follows:
all_gt
spg_all

commented
  • even if you don't have any un-annotated points, the label 0 should be reserved for un-labelled data. Start at 1.
  • replicating the intensity 3 times to mimic RGB is obviously not ideal. dapt here, and change --pc_attribs and --ptn_nfeat_stn to only have one radiometric information.
commented

  • Thank your helpful suggestions！ it works.

  • But the traffic signs line still keeps bad result,only 0.00035 IoU, I used intensity(original intensity of my custom dataset belongs to [0,65535], I change them to [0,1] just likes rgb/255.0) for partition, but it does not work(there are no any traffic sign lines in partition.ply or Geo.ply ),I don't konw why.

  • And May I ask your a simply question.emm..., How to make the GT and Pred at the COMPLETE SAME perspective ,use some software or some open source project? Thank you.

commented
  • Thank your helpful suggestions！ it works.
  • But the traffic signs line still keeps bad result,only 0.00035 IoU, I used intensity(original intensity of my custom dataset belongs to [0,65535], I change them to [0,1] just likes rgb/255.0) for partition, but it does not work(there are no any traffic sign lines in partition.ply or Geo.ply ),I don't konw why.
  • And May I ask your a simply question.emm..., How to make the GT and Pred at the COMPLETE SAME perspective ,use some software or some open source project? Thank you.

Hi, Thanks for your exploration on SPG. May I ask what accuracy do you get in your own dataset.