/second_lite

object detection with lidar point cloud, with real-time performance. This is a simplified version of SECOND project by Yan Yan

Primary LanguagePython

1. Introduction

a lightweight version of SECOND LiDAR object detection

my change:

  • use own customized spconv_lite instead of spconv
  • rewrite/refactor code
  • use a simplfied voxel extractor, but works well.
  • trained with kitti
  • trained with lyft 3d detection
  • add my own understanding and illustrations about the model
  • test on a real testing car, 0.05 sec per Lidar frame, wth GTX 2080

My results

  • KITTI 2011-09-26-0005
  • KITTI 2011 09 26 0023
  • test on a real testing car
    • 3d visualization is my opengl 2.0 practice repo
    • results currently rendered in BEV, yellow is cyclist, defined in KITTI

this is a showcase repo, KITTI training sketch code can be found in https://github.com/masszhou/SECOND_lite_dev

2. SECOND lite Model Architecture

2.1 VFE module

2.2 middle conv3D module

2.3 RPN module

3. How to use

under root of this project

python -m script.predict_kitti predict_pcl_files --pcl_path="{kitti_root}/2011_09_26/2011_09_26_drive_0023_sync/velodyne_points/data/" --image_root_path="{kitti_root}/2011_09_26/2011_09_26_drive_0023_sync/image_02/data/" --save=True

4. Results on KITTI test set

training with 180° FOV labels, inference with 360° FOV

  • green: car
  • blue: van
  • red: pedestrian
  • yellow: cyclist

My results