OpenSigns

Installation

1. Create a conda enviroment with python 3.6.5 and activate it.
   conda create --name sign_det python=3.6.5
conda activate sign_det

2. Install tensorflow 1.13.1 and tensorflow-gpu 1.13.1. It is crucial to install it both with pip and conda.
   pip install tensorflow==1.13.1 --upgrade
conda install tensorflow==1.13.1
pip install tensorflow-gpu==1.13.1 --upgrade
conda install tensorflow-gpu==1.13.1

3. Install keras-retinanet 0.5.1.

   • Clone keras-retinanet repository.
     git clone https://github.com/fizyr/keras-retinanet.git

   • Move to the keras-retinanet folder.
     cd keras-retinanet

   • Create a branch from 0.5.1 and switch to it.
     git checkout -b branch0.5.1 0.5.1

   • Install it.
     pip install . --upgrade

4. Install keras-maskrcnn 0.2.2.

   • Clone keras-maskrcnn repository.
     git clone https://github.com/fizyr/keras-maskrcnn.git

   • Move to the keras-maskrcnn folder.
     cd keras-maskrcnn

   • Create a branch from 0.2.2 and switch to it.
     git checkout -b branch0.2.2 0.2.2

   • Install it.
     pip install . --upgrade

5. Install keras 2.2.5.
   pip install keras==2.2.5 --upgrade

6. Install matplotlib with dependencies.
   pip install matplotlib --upgrade
conda install matplotlib

7. Install exiftool.
   sudo apt-get install exiftool

8. Clone this repository.
   git clone https://github.com/Jozko55/bike_signs_detection.git

9. Download the model separately.
   You may need to adjust PATH_TO_MODEL in script.py.

Usage

Firstly, create a folder where both input pictures and outputs will be stored. The folder must contain two separate folders named inputs and outputs. The subfolder inputs should contain your input pictures. The subfolder outputs should be empty.

Now, adjust accordingly the variable path_to_data in the file script.sh. (By default it is the path to sample_data.)

Finally, run everything.
./script.sh

Results

There will be exactly 4 output files generated for every detected bike sign on an input picture.

• _mask.JPG
• _box.JPG
• _exif.json (generated by exiftool from original picture)
• _info.json (name of detected sign and given score)