Self-Driving Car Engineer Diary — 4

Mon, 30/Jan/2017

Hi. Time to apply our new Deep Learning knowledge to project #2 and classify different traffic signs. This project was an opportunity for exploration and learning aka lots of trial & error!.

I got to explore TensorFlow and various libraries (see table below), different convolutional neural network models, pre-processing images, manipulating n-dimensional arrays and learning how to display results.

Deep Learning Pipeline

TensorFlow framework and other libraries used during the data processing pipeline.

Image Dataset

Sample of the German Traffic Sign dataset.

Pre-processing Image Dataset

Pre-processing with TensorFlow and OpenCV functions.

CNN Model Architecture

Sermanet / LeNet CNN model architecture and spatial reduction transforms.

Training

Results

Looks OK …

… but didn’t ‘generalise’ well when presented with unseen images :(

Compute

Was fun to take my ‘rig’ for a spin!

Next up is Keras and then Transfer Learning.