Unsupervised Logo Detection with Adversarial Domain Adaptation from Synthetic to Real Images

In this chapter, we address the challenges posed by limited training data and domain shift in the field of logo detection. Recent advancements have illustrated the effectiveness of convolutional neural networks (CNNs) trained on simulated or synthetic images for detecting objects in real-world images. Synthesized training images with automatically generated annotations at the object-level offer a promising alternative to the laborious and costly task of bounding box annotation. However, real-world problems limit this assumption and object detectors face domain shift problems, which degrade performance. Knowledge transfer from one domain (synthetic images) to another (real-world images) causes domain shift problems due to the huge differences in data styles and distributions between the source and target domains. This chapter discusses an approach of using only synthesized images for model training and adapting knowledge from unlabelled real-world logo images. We generate synthesized logo images with automatically generated bounding box annotations to facilitate model training. Additionally, to align domain gap synthetic to real-world image, we propose entropy minimization of the mid-level output feature space. Our experiments show that the proposed method improves performance on different logo datasets compared to direct transfer from source to target domain (synthetic-to-real images) without any labeling cost and increasing network parameters.