Mask RCNN Object Detector

A ROS Node for detecting objects using Mask RCNN. The detector is build on the Tensorflow and Keras ecosystem.

Requirements

Install the following dependencies before compiling this package

Install the required deps on the virtualenv

Installation

It is better to use the python virtual environment to install the detector dependencies and note that it only works on Python 3

Install python Virtual Environment

$ sudo pip install virtualenv
$ sudo pip install virtualenvwrapper

Creating Virtual Environment

$ venv_name=mrcnn
$ mkvirtualenv --python=python3 $venv_name

Install the dependencies

$ pip install -r requirements.txt

Downloading the Package

Clone the package to the ROS workspace using git tools

$ git clone https://github.com/iKrishneel/mask_rcnn_ros.git
$ cd mask_rcnn_ros
$ git pull --all
$ git submodule update --init

Compilation

The package is a ROS node and can be complied using the catkin tools

  $ catkin build mask_rcnn_ros
  $ source $HOME/.bashrc

Running

Source to activate the virtualenv containing the installed deps.

Running the node as a publisher

roslaunch object_detector object_detector.launch image:=<image_topic_name> is_service:=false debug:=true

Running the node as a service

roslaunch object_detector object_detector.launch image:=<image_topic_name> is_service:=true debug:=false

Arguments

The following arguments can be set on the roslaunch above.

image: Image topic name
is_service: boolean flag when set launches the node as a ros service.
debug: boolean flag when set shows the detection results in opencv visualization cv.imshow()
detection_threshold: threshold to filter the detection results [0, 1]
model: path to the training keras model file. The model is with .h5 extension.
class_labels: text file containing the mapping of the class label with the class name

Training

The Mask RCNN can be trained as follows on custom dataset.

Dataset Loading

Modify the dataloader script based on your dataset format or use the COCO dataloader. The current dataset loader in this package is different from the COCO loader that comes with the Mask RCNN. It is useful when training Mask RCNN on custom data. Current loader assumes that the dataset directory contains two folders and a textfile:

image/ which contains all the images (eg. image_0001.jpg)
label contains folders (named after the image name eg. image_0001/ corresponding to each image. Inside the subfolder image_0001 are mask images of target objects for learning (eg 1.png, 4.png. The name of the mask image in the subfolder corresponds to the class label of the objects)
class.txt - textfile contains the class names corresponding to each label on a new line(eg: apple 1 [separated by single white space])

Hyperparameter Tunning

Before starting the training, you will need to tune some hyperparameters.

NUM_CLASSES #L53 in the dataset
Learning Params #44 set learning parameters like epochs, layers [which layers of ReNet to train] and LEARNING_RATE (if necessary)

Training Command

$ python mask_rcnn_trainer.py --dataset <path_to_dataset> --model <path_to_model_used_for_initialization>  train