Traffic Sign Classifier is the second project, and the ninth lesson, in the Udacity Self-Driving Car Engineer Nanodegree Program.
In this project, students build and train a deep neural network to classify images from the German Traffic Sign Recognition Benchmark dataset. There are about 40 different types of German traffic signs in the dataset, each 32×32 pixels big. That’s not very big!
Nonetheless, each image is big enough for students to train a convolutional neural network to recognize what type of sign it is, with 95%+ accuracy. That’s close to, or even better than, the accuracy that humans like you and I reach when we classify images by sight.
The lesson starts out with a tour of LeNet, one of the canonical network architectures for image classification. We step through how to implement LeNet in TensorFlow, highlighting data preparation, training and testing, and configuring convolutional, pooling, and fully-connected layers.
We also show students how to spin-up a GPU-enabled EC2 instance from our partners at Amazon Web Services. Thank you to AWS Educate for providing free AWS credits to Udacity students!
At the end of the lesson, students get to apply, tweak, or completely revamp LeNet to train their own classifier. If you want to compare yourself to Yann LeCun, here’s how he did with the same dataset:
Ready to start learning how to build self-driving cars yourself? Great! If you have some experience already, you can apply to our Self-Driving Car Engineer Nanodegree program here, and if you’re just getting started, then we encourage you to enroll in our Intro to Self-Driving Cars Nanodegree program here!