Deep Learning Traffic Sign Recogniser

Zhang et al. A Shallow Network with Combined Pooling for Fast Traffic Sign Recognition

Requirements

• Tensorflow 1.2
• Numpy
• GTSRB Dataset 'gtsrb_dataset.npz'

Replication

By default all our improvements are disabled, so the replication can be run with

python main.py

Improvement

To run our improvemed version the augmented dataset must first be generated

Generating the Augmentated DataSet

To generate the augmented data we use OpenCV to manipulate the training data

Configuring OpenCV

curl https://raw.githubusercontent.com/mitsuhiko/pipsi/master/get-pipsi.py | python
pipsi install backports.shutil_get_terminal_size
pipsi install pew
pip install pipenv --user
pipenv --site-packages  # initialises virtualenv with access to system installation of Tensorflow
pipenv install  # installs packages listed in Pipfile.lock 

Executing the data augmentation

• First enter the pipenv environment: pipenv shell
• Generate augmented data: python augment_data.py generate

Optional

Viewing augmented data python augment_data.py show 1000 Where 1000 is the index of the image you want to view variants of. Environment variable SCIPY_PIL_IMAGE_VIEWER must be set to an image viewer.

Running the improved version

python main.py --multi-scale --crelu --use-augmented-data --dropout-keep-rate=0.7 --max-pools

Architecture

Layers

Layer	Name	Kernel	Stride	Padding	Output
1	Input	-	-	-	(32,32,3)
2	Convolution/Relu	5x5	1	2	(32,32,32)
3	Average Pool	3x3	2	[0 1 0 1]	(32,16,16)
4	Convolution/Relu	5x5	1	2	(32,16,16)
5	Average Pool	3x3	2	[0 1 0 1]	(32,8,8)
6	Convolution/Relu	5x5	1	2	(64,8,8)
7	Max Pool	3x3	2	[0 1 0 1]	(64,4,4)
8	Convolution/Relu	4x4	1	2	(64,1,1)
9	Fully Connected	1x1	1	0	(64)
10	Softmax Loss				(43)