Dynamic Laser Scanning Dataset for Multi-view Incremental Segmentation

Supplementary material (dynamic laser scanning dataset) for the RAL paper Multi-view Incremental Segmentation of 3D Point Clouds for Mobile Robots.

Prerequisites

numpy
scipy
scikit-learn
tensorflow
ROS
h5py

Ray Tracing

Obtain the S3DIS dataset from here, which is from the paper 3D Semantic Parsing of Large-Scale Indoor Spaces by Armeni et al.. Run the following code to generate a ROS bag file as a result of the scan simulation.

#combine all rooms from Area 3 to a single HDF5 file
python building_parser_combined.py --area 3

#perform ray-tracing to generate a bag file
python raytrace_dynamic.py --area 3

Scan Data

The point cloud data contains 3D coordinates (XYZ), color (RGB), object instance ID (O), and class ID (C) for each scan point.
The bag file structure is as follows:

Topic	Data Type	Description
laser_cloud_surround	sensors_msgs/PointCloud2	Array of (X,Y,Z,R,G,B,O,C) tuples
slam_out_pose	geometry_msgs/PoseStamped	Robot pose at each scan point
trajectory	nav_msgs/Path	Array of sequences of robot poses

Usage

#optional: train a model from scratch
#select MCPNet as the network architecture (other options are pointnet,pointnet2,voxnet,sgpn)
#select Area 3 as validation set (Areas 1,2,4,5,6 as training set)
#model will be saved in models/mcpnet_model3.ckpt
python train.py --net mcpnet --dataset s3dis --train-area 1,2,4,5,6 --val-area 3

#optional: train a model for the outdoor dataset
python train.py --net mcpnet --dataset outdoor --train-area 5_0,7_0,9_1,10_2,10_6 --val-area 16_4

#start the ROS node for incremental segmentation
#select Area 3 as the validation dataset
#select MCPNet as the network architecture
#use the flag --color to publish original color point cloud scans
#use the flag --cluster to publish clustering results
#use the flag --classify to publish classification results
python inc_seg.py --net mcpnet --area 3 --dataset s3dis

#use RViz as a visualization tool
rviz -d inc_seg.rviz

#publish the laser scan data from a ROS bag file
rosbag play data/s3dis_3.bag

Evaluation

#calculate evaluation metrics after replaying from ROS bag files
for i in 1 2 3 4 5 6
do
    python -u inc_seg_replay.py --net mcpnet --area $i --dataset s3dis --save >> results/result_mcpnet.txt
done

Evaluation for offline methods

#train an offline PointNet model from scratch
for i in 1 2 3 4 5 6
do
    python -u train_offline.py --net pointnet --area $i > results/log_train_offline_pointnet_area$i.txt
done

#calculate evaluation metrics using offline PointNet (with room segmentation)
for i in 1 2 3 4 5 6
do
    python -u test_offline.py --net pointnet --area $i >> results/result_offline_room_pointnet.txt
done

#calculate evaluation metrics using offline PointNet
for i in 1 2 3 4 5 6
do
    python -u offline_seg.py --net pointnet --area $i >> results/result_offline_pointnet.txt
done

#calculate evaluation metrics using semi-offline PointNet
for i in 1 2 3 4 5 6
do
    python -u semi_offline_seg.py --mode space --area $i >> results/result_semi_offline_space.txt
done

Non-learning baseline methods

#online segmentation using normal vectors
python baseline_seg.py --mode normals --dataset outdoor --area 16_4

Reference

@article{Chen2019,
    author = {Chen,Jingdao and Cho, Yong K. and Kira, Zsolt},
    title = {Multi-view Incremental Segmentation of 3D Point Clouds for Mobile Robots},
    journal = {IEEE Robotics and Automation Letters},
    year = {2019},
}