Filtering local features for logo detection and localization in sports videos

This paper presents a system for the detection and localization of multiple instances of trademark logos in sports videos. It is based on SIFT features and considers the local geometry of neighboring features in order to differentiate between different logos with ambiguous local features such as text-based logos. In contrast to other approaches, we do not rely on a training phase and therefore no labeled data with annotated or absent logos is needed. The focus of the detection approach is on images of sports videos which suffer from compression artifacts, motion blur, small object sizes, occlusion and several other artifacts. Results are presented on video images of a soccer game containing logos on different advertising media.