Traffic Sign classifier using CNN and TensorFlow
The goal of this project was to implement a Convolutional Neural Network to detect German Traffic Sign using TensorFlow. The entire model was trained a ubuntu instance on Google Cloud Platform (GCP) with a NVIDIA Tesla K80 GPU.
The data set used is the German Traffic Sign data set that can be dowloaded from here. German Trafffic Sign Dataset is a set of 51,839 labeled images of 43 different German traffic signs. It comes in two separate sets. A set of 39,209 images for training and another set of 12,630 in order to test the accuracy of our trained network. A pickled version of the dataset can be downloaded from here.
The CNN architecture used was LeNet architecture with 2 CONV layers and 3 fully connected layers. Activation function was ReLU. Training was done for 30 epochs with a batch size of 128 using Adam Optimizer. I was able to obtain a validation accuracy of 96.4% and test accuracy of 91.8%