search

Natsu-Akatsuki / RangeNetTrt8

tensorrt8 && cuda && libtorch implementation of rangenet++
MIT License
44 stars 9 forks source link

modify the code applicable to Ouster64 data #19

Closed G12311231 closed 2 months ago

G12311231 commented 2 months ago

Hello， Thank you very much for the open source！

I tried to directly use the pre-trained model provided by lidar-bonnetal to segment the data collected by ouster64. The segmentation results were very strange. How can I modify the code to make it applicable to ouster64 data?

Natsu-Akatsuki commented 2 months ago

I’m afraid not. The pretrained model is specifically tailored for the Velodyne-64.

G12311231 commented 2 months ago

Thank you for your answer! I want to know if it is because the rangenet++ method converts a three-dimensional point cloud into a two-dimensional one based on a distance image, so different types of sensors cannot use the model trained by the author based on vlp-64. And If I use the same sensor vlp-64 to collect data on campus, would like to know if this dataset can use the parameters of the pre-trained model?

Natsu-Akatsuki commented 2 months ago
  1. Yes. This is one reason. It is welcome to implement the projection method (3D -> 2D) to adapt for Ouster-64 and it may work. Another concern is the domain gap between the two LiDAR, which may lead to decrease in accuracy.
  2. Look so good. However, as I have not personally conducted the tests, I cannot provide you with a definitive answer with 100% confidence.
G12311231 commented 2 months ago
  1. Yes. This is one reason. It is welcome to implement the projection method (3D -> 2D) to adapt for Ouster-64 and it may work. Another concern is the domain gap between the two LiDAR, which may lead to decrease in accuracy.
  2. Look so good. However, as I have not personally conducted the tests, I cannot provide you with a definitive answer with 100% confidence.

I understand, and thank you very much for your reply!