The aim of this project is to improve an existing solution of object detection, based on the excercise of the official Duckietown project documentation: Unit B-5.
We expect to enhance the performance of the network by using Intel Movidius Accelerator stick and a modified network architecture.
For training we are using the datasets given in the baseline with minor modifications in the separation of the three main sets for training, validation and testing.
The used data originates from the darknet github repository. The dataset contains 420 labeled images where the labeling is attached in a separate text file. In order to enlarge the dataset, we use Augmentor for generating images with slightly different properties. The tool is able to generate the images on the fly, implementing a keras generator, so storing the generated images is not neccessary. For example import, see the training notebook.
We used the attached template LaTeX file. The docs can be found under the docs folder. The compilation can be made using the following method.
cd docs
docker run -v `pwd`:/home -it tianon/latex
cd /home
pdflatex DuckTales_assignment.tex