Xiang Ruan

The scalability and applicability of fully supervised logo detectors are severely limited due to their strong reliance on object-level labeled data for training, which is tedious, expensive, and time-consuming. In contrast, weakly supervised object localization techniques rely on image-level annotations, resulting in insufficient detection efficiency. Some researchers have explored the use of synthetic images to address the challenges of object-level data annotation. However, models trained only on synthetic images often perform subpar on real-world images due to the significant domain shift in data characteristics between the training and testing phases. Unsupervised domain adaptation remains a challenging problem for various tasks, including logo detection. To address these issues, we present an improved weakly supervised logo detection approach that utilizes real-world images with image-level annotations, combined with synthetic images featuring automatically generated annotations. Our detection framework employs image-level supervision for both the target (real-world images) and source domains (synthetic images), leveraging the object localization capabilities of CNNs, which has been proven to be crucial in aligning domains. To effectively reduce the domain gap between synthetic and real-world images, we introduce a multi-class discriminator CNN network to learn an aligned semantic probability distribution space that adapts the target domain characteristics. Additionally, we propose a regularization technique that identifies hard object samples in the target domain to enhance the extracted features in a generalized space. Experimental results demonstrate that our method outperforms existing domain adaptation methods for logo detection task.