mmdetection-ros

This is a ROS package for 2D object detection, which simply utilizes the toolbox MMDetection of OpenMMLab.

Requirements

• ROS Melodic
• Python 3.6+, PyTorch 1.3+, CUDA 9.2+ and MMCV

Installation

0. Clone all needed packages: this package, mmdetection and vision_msgs

   cd {YOUR WORKSPACE}/src
   git clone https://github.com/afengisme11/mmdetection_ros.git
   git clone https://github.com/ros-perception/vision_msgs.git
   cd mmdetection_ros
   git clone https://github.com/open-mmlab/mmdetection.git

1. MMDetection is a toolbox for 2D object detection. For using it I used a virtual environment in python3

   sudo apt-get install python-pip python-dev python-virtualenv
   virtualenv --system-site-packages --python=python3 ~/mmdet
   source ~/mmdet/bin/activate
   pip --version # check the pip version

2. In this created environment we install all the requirements. Note that the following instructions for mmdetection are almost the same to the official one here, so you may also firstly have a look there if you have problems.

   pip install mmdet

3. Then install PyTorch and torchvision following the official instructions here or for previous version here, e.g., I use CUDA 10.2 and install PyTorch 1.7.1, so I run this command:

   pip install torch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2

4. Install mmcv-full.

   pip install mmcv-full -f https://download.openmmlab.com/mmcv/dist/{cu_version}/{torch_version}/index.html

   Check the table here. Please replace {cu_version} and {torch_version} in the url to your desired one. For example, I use the following command:

   pip install mmcv-full -f https://download.openmmlab.com/mmcv/dist/cu102/torch1.7.0/index.html

5. Install build requirements and then install MMDetection.

   cd ./mmdetection
   pip install -r requirements/build.txt
   pip install -v -e .

6. Install rospkg.

   pip install rospkg

7. Put your model in the scripts folder, and modify the model path and config path in the mmdetector.py. The default one now is a Faster R-CNN I trained on my custom dataset to detect 5 products in the robot TIAGo context. You can also get it here if you want

ROS Interfaces

params

• ~publish_rate: the debug image publish rate. default: 1hz
• ~is_service: whether or not to use service instead of subscribe-publish
• ~visualization: whether or not to show the debug image

topics

• ~debug_image: publish the debug image
• ~objects: publish the inference result, containing the information of detected objects
• ~image: subscribe the input image. The default one is ''/xtion/rgb/image_raw' for TIAGo

Example

If you are using the robot TIAGo, you could

source ~/mmdet/bin/activate # in the python3 virtual env
roslaunch mmdetection_ros mmdetector.launch

TIAGo will continously do the inference on the camera image and publish the detection results