Summary

In this chapter, we focused on pre-trained neural networks, and how we can leverage them for our purposes. We combined two neural networks to detect pedestrians, vehicles, and traffic lights, including their color. We first discussed how to use Carla to collect images, and then we discovered SSD, a powerful neural network that stands out for its capacity to detect not only objects, but also their position in an image. We also saw the TensorFlow detection model zoo and how to use Keras to download the desired version of SSD, trained on a dataset called COCO.

In the second part of the chapter, we discussed a powerful technique called transfer learning, and we studied some of the solutions of a neural network called Inception, which we trained on our dataset using transfer learning, to be able to detect the colors of traffic lights. In the process, we also talked about ImageNet, and we saw how achieving 100% validation accuracy was misleading, and as a result, we had to reduce...