arekmula / ros_handler_detector

ROS node for handler detection of articulated objects.

ros_handler_detector

About

               

The goal of the project is to build a ROS node that would be responsible for detecting handlers of articulated objects such as cabinets, wardrobes, or lockers. The module uses a neural network to perform the task and utilizes CenterNet Resnet50 V1 architecture. The dataset used for training, evaluation, and testing is available here

This module is part of my master thesis "Point cloud-based model of the scene enhanced with information about articulated objects" and works best with the other three modules that can be found here:

• Front detector
• Rotational joint detector
• Articulated objects scene builder

Results

• mAP@IoU=.50 -> 0.928
• mAP@IoU=.75 -> 0.473
• mAP@IoU=0.50:0.95 -> 0.503

Prerequisities

• Ubuntu 20.04
• ROS Noetic
• Tensorflow 2
• Python 3.8

Installation

mkdir -p caktin_ws/src
cd catkin_ws
catkin_make
cd src
git clone https://github.com/arekmula/ros_handler_detector
cd ros_handler_detector/src
protoc object_detection/protos/*.proto --python_out=.
cd ~/catkin_ws
rosdep install --from-path src/ -y -i
catkin_make

Run

• Setup path to your model directory and label map:

rosparam set model_dir "path/to/model"
rosparam set label_map_path "path/to/labelmap"

• Setup RGB image (640x480) topic:

rosparam set rgb_image_topic "image/topic"

• Determine if visualization image should be published

rosparam set visualize_handler_prediction True/False

• Run with

rosrun handler_detector handler_detector.py

• Change detection threshold if you want

rosparam set handler_prediction_threshold 0.7