Some work on the comma.ai self driving cars dataset.

The proposed architecture is a slightly modified version of the PilotNet published by Nvidia. The following improvements are proposed to ameliorate training and generalization capabilities.

1. Batch normalization after each convolutive layer.
2. Activation function type 'elu'. For more information read this article.
3. Dropout before the output layer.

Some nice functions to annotate input images and render .AVI videos with the autopilot predictions are provided. These functions were used to generate the gif showed above.

1. Annotate input images with vehicle's CAN readings.
2. Generate video with the autopilot predictions and target angles.

Creative Commons Attribution-NonCommercial-ShareAlike 3.0 License