Mitigating Domain Shift in Logo Detection: An Adversarial Learning-Based Approach

This chapter discusses a domain adaptation-based technique aimed at mitigating the domain-shift challenge in logo detection. In real-world applications, we face a domain shift problem between the training data (source domain) and test data (target data) resulting in reduction of performance. The domain gab or domain shift problem is caused by the difference in feature distributions of training and test data. In practical scenarios, deploying trained models to detect objects in different images at various platforms, which may possess different data feature distributions and styles compared to the training data, often results in decreased performance. As a result, re-training of models with user-specific annotated training data is required, which is time-consuming, laborious and expensive. It is impractical for real-world applications. To address this issue, this study introduces domain adaptation-based technique to train detection framework, aligning networks across datasets from different logo datasets. The proposed method uses unlabelled data samples from target domain alongside labelled source domain data during model training to generalize the detection framework. To bridge the gap between different domains we exploit the adversarial-domain adaptation learning. This is a pragmatic way of dealing with the domain-shift problem using an anchorfree object detector.