Traffic-Sign-Recognition

Traffic sign classification on the German Traffic Sign Recognition Dataset.

Requirements

The project uses Python3.6 with Keras (Tensorflow backend).

The GTSRB dataset can be downloaded from this link.

Usage

Training

To train a model, run the following command.

python train.py [-h] [-d BASE_DIR] [-m MODEL] [-e EPOCHS] [-b BATCH_SIZE]
                [--lr LR] [--num_classes NUM_CLASSES] [-c CHECKPOINT]

optional arguments:
  -h, --help            show this help message and exit
  -d BASE_DIR, --base_dir BASE_DIR
                        path to dataset (default: None)
  -m MODEL, --model MODEL
                        Model to use. Choose between LeNet_baseline,
                        LeNet_modified, AlexNet, VGG16, VGG19 and ResNet18
                        (default: LeNet_baseline)
  -e EPOCHS, --epochs EPOCHS
                        number of epochs to run (default: 25)
  -b BATCH_SIZE, --batch_size BATCH_SIZE
                        batch size (default: 1)
  --lr LR               learning rate (default: 1e-3)
  --num_classes NUM_CLASSES
                        Number of classes (default: 43 for GTSRB)
  -c CHECKPOINT, --checkpoint CHECKPOINT
                        path to checkpoint directory

Architectures currently supported are: LeNet-5, AlexNet, VGG16, VGG19, ResNet18.

Testing

To evaluate performance of a trained model, run the following command.

python test.py [-h] [-m MODEL] [--pretrained_model PRETRAINED_MODEL]
               [-d BASE_DIR] [--num_classes NUM_CLASSES]

optional arguments:
  -h, --help            show this help message and exit
  -m MODEL, --model MODEL
                        model to use (default: None)
  --pretrained_model PRETRAINED_MODEL
                        path to pretrained model (default: None)
  -d BASE_DIR, --base_dir BASE_DIR
                        path to dataset (default: None)
  --num_classes NUM_CLASSES
                        Number of classes (43 for GTSRB)

Performance

Due to computational constraints, only LeNet could be trained. Other architectures were tested for correctness by training for 1 epoch.

Model	Val accuracy	Test accuracy
LeNet baseline	98.113%	91.37%
LeNet modified	98.636%	94.36%

The following hyperparameters were used: