Point-cloud-compression-by-RNN

• PyTorch implementation of our work published in ICRA2019
• Support data from Velodyne HDL64E(S2,S3), HDL32E and VLP16 sensor

Requirements

• Linux (tested on Ubuntu 16.04)
• Python 3.5+
• PyTorch 0.3.1+
• ROS
• Opencv

Training your own network

Install

Build tools library (to convert LiDAR packet data into 2D matrix or reconstruct packet data back) by running the following command:

cd tools/src/
catkin_init_workspace
cd ../
catkin_make -DCMAKE_BUILD_TYPE=Release

Build training data

Download training data from here or use you own data.
Make sure rosbags have /velodyne_packets topic and copy them to /training_data. Run

cd training_data
./extract.sh

to build the trainging data.
Each frame of point cloud from LiDAR is decomposed into a 2D intensity matrix (.png unrelated to our compression) and a rotation angle vector (.txt can be further compressed into a few bits by run length code) and a distance matrix (.png our main target and input of network)

Training

cd network
python training.py  

Examples

Here we provide an example of compression and decompression. We provide 32 frames' LiDAR data from Velodyne HDL32E sensor in network/Evaluation/raw_decompose

1. Compression and decompression

./enc_dec.sh

All frames are compressed using max number (32) of interval, while decompressed from 10 to max number of interval.

2. Reconstruction to point cloud (.pcd) file

./rearrange.sh
./rearr2recon.sh 

Acknowledgement

• 1zb/pytorch-image-comp-rnn: Our network's code references this work