Chapter 6

1. What is the particularity of autoencoders?

Autoencoders are encoders-decoders whose inputs and targets are the same. Their goal is to properly encode and then decode images without impacting their quality, despite their bottleneck (that is, their latent space of lower dimensionality). 

2. Which classification architecture are fully convolutional networks (FCNs) based on?

FCNs use VGG-16 as the feature extractor.

3. How can a semantic segmentation model be trained so that it does not ignore small classes?

Per-class weighing can be applied to the cross-entropy loss, thereby penalizing more heavy pixels from smaller classes that are misclassified. Losses that are not affected by the classes' proportions can also be used instead, such as Dice.