/Multi_view_Cylinder_Net

range map based lidai odomentry

Primary LanguagePython

This is a little demo of deep-learning based odomentry write by myself. The demo has tested on the Kitti Dasaset.

Multi_view_Cylinder_Net

An end-to-end deep-learning-based lidar odometry.

In order to ensure the speed of odometry, cylinder projection of lidar points is adopted.

Pipeline

NetWork

image-20210121232134936

Illustration of multi-view projection and segmentation

image-20210121232308753

Run

Dependencies

python>3.6
pytorch>=1.5
cupy
tqdm
shutil
matplotlib
glob
argparse

Train

Firstly, change the config information in args.py.

Then run,

python train.py

Test

Change the config information in args.py and run,

python test.py

Results on Kitti

  • 01

    image-20210121232500701

  • 03

    image-20210121232530428

  • 04

    image-20210121232546114

  • 07

    image-20210121232620545

  • 08:

    image-20210121232635286

  • 09:

    image-20210121232648109

  • 10:

    image-20210121232708299

Note: sequence 00-08 for training, 09 and 10 for testing.