LogoMix: A Data Augmentation Technique for Object Detection Applied to Logo Recognition
Daniel Fuertes, Carlos Roberto Del-Blanco, Fernando Jaureguizar, Narciso Giarcia · 2022
Automatic logo detection on images is a challenging task in computer vision with many applications in social media platforms. Recent works use generic deep-learning based object detectors, but they require huge databases for optimal recognition performance, which are very costly to obtain, especially in logo detection applications where the number of classes can be very high. For this reason, LogoMix, a new data augmentation method for object detection tasks, is proposed in this paper. It creates new sample images by combining logo entities with different degrees of overlapping. A state-of-the-art object detector trained with LogoMix has been evaluated on two popular logo datasets, QMUL-OpenLogo and FlickrLogos-32, and compared with other state-of-the-art works to prove its high performance.