This repository use multiple predefine model in keras, and try to fine tune them on a traffic sign classification task.

• Keras
• numpy
• matplotlib

This script expect a dataset with the following structure:

Dataset
├── Class1
│   ├── image1.jpg
│   └── image2.jpg
├── Class2
│   ├── image1.jpg
│   └── image2.jpg
└── Class3
    ├── image1.jpg
    └── image2.jpg

python3 train.py <super-class-to-use> <dataset-path>

Where super-class-to-use is one of the first level key of classes dict in train.py.

A lot of different parameters can be tuned on the command line call, to have a complete description of them, please run:

python3 train.py -h

The predefined super classes of classes dict are MUTCD code of US signs for super classes Rectangular, Diamond and Zebra. For RedRoundSign it is the class names as used in the TT100k traffic sign dataset.

An other project, based on Siamese network, showed very promising results to tackle the problem of sign classification (low data sample for some class, lot of classes). This project code is available at https://github.com/DL-project-Fall2019/Siamese-traffic-signs

Nicolas Six