Few-shot Logo Recognition in the Wild

Abstract

Brand logo recognition can be viewed as identification and classification task. It finds many usages such as market discovery, target advertising, etc. The number of logos growth every year and logo itself can appear in vast variety of contexts, therefore we propose a two-step few-shot framework. We describe a novel combination of universal logo detector and few-shot classifier. The logo detector is based on YOLOv5 and is used to find the areas on the image where logos are located. With this state-of-the-art single-stage object detector we achieved higher precision than similar double-stage solutions. To classify detected logos we propose few-shot classifier which consists of ensemble of pretrained feature extractors and fine-tuned head.