In deep learning applications large annotated datasets are considered necessary for application development and improved model performance. This work aims to investigate the validity of this assumption when enlarging a given dataset, by secondary data, with a certain domain discrepancy. The paradigm for this evaluation is a vehicle reidentification system for city-scale multi-camera settings. In city-scale multi-camera settings, the field of view of the sensors are fixed, introducing a major domain discrepancy between different datasets. This work shows that the domain of training samples heavily influences the learned feature space embedding and thus leads to a domain-specific performance. We explore how different objective functions and transfer learning approaches cope with a domain discrepancy in the training data. Concluding, the general assumption “Data is of the essence” has to be refined. With respect to feature space embeddings, our findings propose, beyond data “Domain is of the essence”.