Domain generalization learns from one or multiple source domains. It aims to extract a domain-invariant model that can be employed in an unknown target domain. During network training, the models may become overly reliant on the distribution of the training datasets, ultimately leading to an inability to capture the appropriate patterns present in the unseen data accurately.

Framework of the proposed domain generalization method. (Image by the research group)

To alleviate these impacts, Machine Intelligence Group from Shenyang Institute of Automation(SIA), the Chinese Academy of Sciences, proposed a domain generalization method based on weighted label smoothing regularization, relying on the consistency of relative semantic relationships between categories in different data domains.

The progress was publihsed in Knowledge-Based Systems on 30 Jan.

This approach incorporates inter-class supervision through weighted label smoothing regularization, enabling the model to focus more on the relative relationships between categories rather than their absolute differences. This can help the model learn more generalizable features, alleviating overfitting and improving the model’s ability to generalize to unknown new data. Additionally, the weight adaptation technique seeks a balance between transferability and discriminability, stabilizing network updates.

The research provides new ideas for addressing visual perception challenges under varying lighting and weather conditions in the real world.