Deprecated: Development moved to ThundeRatz/ros_yolo2, which provides much faster training and testing.

ROS DetectNet integration

Integrates DetectNet models (eg. generated with DIGITS) with ROS, generating messages with positions of detections.

Requirements

• ROS (tested on kinetic)

• Caffe with GPU support

• CUDA

Installation

This repository should be cloned to the src directory of your workspace. Alternatively, the following .rosinstall entry adds this repository for usage with wstool:

- git:
    local-name: detectnet
    uri: https://github.com/ThundeRatz/ros_detectnet.git

Setup

Copy your model deploy.prototxt and a snapshot of the network's weights as snapshot.caffemodel to the data folder.

ROS Topics

Raw images are received in the image topic and results are published as sensor_msgs/RegionOfInterest messages at detectnet.

For example, the following launch file can be used to map a images from usb_cam to detectnet:

<launch>
  <group ns="vision">
    <remap from="image" to="usb_cam/image_raw" />
    <node pkg="usb_cam" name="usb_cam" type="usb_cam_node" required="true" />
    <node pkg="detectnet" name="detectnet" type="detectnet_node" required="true" />
  </group>
</launch>

Contributors

Written by the ThundeRatz robotics team.

Caffe code is based on ros_caffe, which provides ROS integration with Caffe for image classification.