nearby nano-drone; in this context, we propose a data augmentation approach based on
synthetic background substitution to learn a lightweight CNN model from a small real-world
training set. Experimental results on data from two different labs proves that the approach
improves generalization to unseen environments.