PWC-Net ROS

ROS package to estimate optical flow by PWC-Net.

This uses model definition and trained model from official implementation by Caffe. The model is fine-tuned by Sintel, KITTI, and HD1K dataset.

Requirements

• Nvidia GPU

• Docker

• Docker Compose

• nvidia-container-toolkit and nvidia-docker2

  nvidia-docker2 is deprecated but it is needed for Docker Compose with GPU

Run sample

$ git clone https://github.com/fujimo-t/pwc_net_ros.git
$ cd pwc_net_ros/docker
$ xhost +local:root # To use GUI, see http://wiki.ros.org/docker/Tutorials/GUI#The_simple_way
$ docker-compose up

Then containers is launched:

• ROS master
• rqt
• terminal
  • To run command and ROS nodes

To test pwc_net_ros, execute follow command in the container terminal:

$ roslaunch pwc_net sample.launch

libpwc_net

Use this library to estimate optical flow. You can know how to use it by reading source code of sample_node

sample_node

A node estimates dense optical flow from image topic.

Subscribed topic

• image (sensor_msgs/Image)

  Input image should be remapped. Optical flow is estimated between latest image and it's previous image.

Published topic

• optical_flow (sensor_msgs/Image)

  Estimated optical flow. encoding is 32FC2 (32bit float, 2 channels). First channel is optical flow's x-axis component, second is y-axis.

• visualized_optical_flow (sensor_msgs/Image)

  Visualized optical flow as BGR image to see on normal image viewer such as RViz.

Parameters

• ~image_transport (string, default: "raw")

  Transport used for the image stream. See image_transport.

License

See LICENSE.

This repository doesn't directly contain PWC-Net code but used with it. See LICENSE.md about PWC-Net's license.