Domain generalization and synthetic data in object detection: the enabler, the probe, and the gap
Abstract
This paper reviews the role of synthetic data in enhancing domain generalization for object detection models, which often struggle with performance under distribution shifts.
Reality Card
The paper argues that synthetic data can significantly improve the robustness of object detection models against distribution shifts by providing diversification and alignment strategies.
The paper identifies the synthetic-to-real gap as a major challenge in deploying models trained on synthetic data to real-world scenarios.
Current domain generalization approaches for object detection do not adequately address the complexities of localization and classification under domain shifts.
Paper to code
Verified implementation resources so builders can test the paper’s claims instead of stopping at the abstract.