Brand logo detection in social media images [Sosyal medya görsellerinde marka logosu tespiti]

Official URL: https://dx.doi.org/10.1109/SIU71813.2026.11636669

Although logos in social media images are often not the focal in the post, they provide critical signals for measuring brand visibility. Existing open logo benchmarks partially reflect social-media-specific challenges such as highly diverse contexts, cluttered backgrounds, and small or partially occluded logos, which can limit model generalization in real deployments. In this work, we present a social-media-focused logo detection dataset covering 25 widely used brands in Türkiye. The data collection and annotation pipeline was designed with quality assurance, including analyses of annotator behavior, session dynamics, and label consistency. The resulting dataset contains 5,087 images and 9,731 logo instances. Our main objective is to quantify the impact of training data source on detection performance. We therefore compare a model trained on our social media dataset against a model trained on a Flickr dataset, and evaluate each model on the other dataset's evaluation split. Cross-dataset results show that the social-media-trained model achieves stronger in-domain performance and better out-of-domain generalization, indicating higher robustness for social media logo detection.