Skip to content

Latest commit

 

History

History
62 lines (42 loc) · 2.32 KB

README.md

File metadata and controls

62 lines (42 loc) · 2.32 KB
Raw

Semantic 3D mapping

This package builds a semantic 3D grid map using stereo or RGBD input. It is suitable for large scale online mapping. The grid semantic label is optimized through hierarchical CRF.

Authors: Shichao Yang, Yulan Huang

Related Paper:

  • Semantic 3D Occupancy Mapping through Efficient High Order CRFs, IROS 2017, S. Yang, Y. Huang, S. Scherer PDF

Installation

Prerequisites

This code contains several ros packages. We have test in ROS indigo/kinect. Create or use existing a ros workspace.

mkdir -p ~/mapping_3d/src
cd ~/mapping_3d/src
catkin_init_workspace
git clone [email protected]:shichaoy/semantic_3d_mapping.git

Download Data

cd semantic_3d_mapping/grid_sensor
sh download_data.sh

if the download link breaks, please download here and follow sh file to process it. Will fix it later.

If wget not installed, sudo apt-get install wget

Compile

cd ~/mapping_3d
catkin_make

Running

source devel/setup.bash
roslaunch grid_sensor grid_imgs.launch

You will see point cloud in Rviz. It also projects 3D grid onto 2D image for evaluation, stored at grid_sensor/dataset/crf_3d_reproj .

Notes

  1. Some mode parameters can be changed in grid_sensor/params/kitti_crf_3d.yaml and grid_imgs.launch

    if use_crf_optimize = false, no 3D CRF optimization, 2D label is directly transferred to 3D.

    if use_crf_optimize = true, use_high_order = false dense 3D CRF optimization runs.

    if use_crf_optimize = true, use_high_order = true High order 3D CRF optimization runs. Superpixel data needs to be provided

  2. This package only contains grid mapping, all the pre-processing steps are not included. See preprocess_data/README.md for details.

    Elas is used for computing dense disparity/depth. Dilation CNN is used for 2D semantic segmentation.

    SLIC is used for generating superpixel. ORB SLAM is used to estimate camera pose.

  3. If Grid sensor memory is not initialized properly, delete /dev/shm/sem.shared_grid_map...

  4. Our used ground truth image annotations are in preprocess_data/gt_label/ Refer to paper experiments for more details.