nakajimakou1/yolact_ros

Use YOLACT in ROS

yolact_ros

This is a ROS package interfacing real-time object detector YOLACT (https://github.com/dbolya/yolact) which provides the masks of the detected objects. A ROS node subscribes bgr image and publishes bgr image, class labels, prediction socres, bounding boxes and masks. This ROS package uses the same trained weights and configurations as in YOLACT.

Pre-request

• ROS with Python 3.5
• Pytorch and torch vision

Compatability

Tested on Ubuntu 16.04 + ROS Kinetic + Python 3.5
To use cv_bridge with Python 3 in ROS Kinetic, build cv_bridge from source in a catkin workspace, remember to set python library to Python 3.

Citation

@misc{leiYolactROS2019,
  title = {{ROS} Interface for Real-Time Object Detector {YOLACT}}, 
  author = {Lei Shi},
  url = {https://github.com/leishi07/yolact_ros},
  year = {2019},
}

Usage

1. Launching

roslaunch yolact_ros yolact_ros.launch

2. Parameters

Parameters can be changed in launch/yolcat_ros.launch

• sub_img_topic_name: ROS image topic to subscribe

• model_path: path to save model and name of the model, weights can be downloaded at YOLACT site (https://github.com/dbolya/yolact)

• cuda: use GPU acceleration or not

• use_fast_nms: use fast NMS or not

• threshold: threshold for detection

• display_cv: Display cv image or not

• top_k: get top k objects, you will also only get k objects in published ros topics

• yolact_config: set configuration for Yolact

  "yolact_config"	Model
  yolact_base_config	yolact_base_54_800000.pth
  yolact_darknet53_config	yolact_darknet53_54_800000.pth
  yolact_resnet50_config	yolact_resnet50_54_800000.pth
  yolact_im700_config	yolact_im700_54_800000.pth

3. Published Topics

• sensor_msgs/Image: "/yolact_image"
  Publishes ROS image topic.

• yolact_ros/yolact_objects: "/yolact_objects"
  Publishes detected objects topic. This is a custom message type.

  yolact_ros/yolact_objects:

  Header header
  Object[] object
  

  yolact_ros/Object:

  string Class
  float64 score
  int64 x1
  int64 y1
  int64 x2
  int64 y2
  sensor_msgs/Image mask
  

  x1 and y1 is the top-left point of the bounding box x2 and y2 is the bottom-right one.