Domain generalization and synthetic data in object detection: the enabler, the probe, and the gap
The paper investigates Domain Generalization (DG) in object detection using synthetic data from three angles: as a tool for diversification and alignment strategies, as a probe to find failure modes, and as a source of the synthetic-to-real gap. It shows how these methods enable better generalization by creating diverse training data, learning domain-invariant features, testing robustness under shifts, and bridging the gap between simulation and reality.