U15-Logos: Unconstrained Logo Dataset with Evaluation by Deep learning Methods

Detecting the appearance of logos in images is applied to many applications, such as brand recognition for marketing analysis and intellectual property protection. It is a challenging problem in the computer vision field, especially in unconstrained images with a complex background, difference in sizes, a number of logos per image, contextual diversity with uncontrolled illumination, and low-resolution. In this work, we tackle this problem by first introducing a newly collected datasets containing 15 well-known brands, namely, U15-Logos (unconstrained logos) with 15000 images. In order to ease experimental evaluation and consider model adaptation, we use our standard dataset called F15-Logos (fundamental logos) with 15042 images for a crossing evaluation between U15-Logos and F15-Logos. Finally, we provide an evaluation of the state of the art models based on Deep learning including YOLO, Faster RCNN, Mask RCNN, RetinaNet in order to illustrate how well these models overcome unconstrained conditions. Comparative evaluations demonstrate the best performance on U15-Logos is about 89,4%.