Planar Trademark and Logo Retrieval

Sport advertising has become an important business increasingly raising the interest of an efficient analysis. To reduce the manual workload, in this work we present an automatic specific trademark and logo recognition system overcoming typical problems of existing (mostly SIFT-based) approaches. In particular, we need to cope with relatively small or correlated trademarks, severe background clutter, and huge perspective variations. This is realized by introducing a concept to increase the perspective invariance, a sophisticated verification, a guided matching phase that is able to deal with a vast number of outliers, and the use of an additional complementary interest region detector with multi-resolution shape description. To show the benefits of the approach, we demonstrate it for a representative real-world test set consisting of images of the EURO 2008 final game. The results clearly show that using the proposed method existing approaches can be outperformed in terms of accuracy and recall.