Framework for Fine-grained Recognition of Retail Products from a Single Exemplar

Abstract

We propose a framework which allows one-shot fine-grained recognition of retail products in a real store from clean images used in e-commerce websites. We apply a metric learning approach to train the one-shot recognition model. To learn a suitable metric space for classification, we construct a data collection system which efficiently captures a large variety of products from various viewpoints under controllable lighting conditions. This dataset plays a role of an intermediate domain between the clean images and real stores. To expand applicable area of the intermediate domain, we use a domain generalization technique. In addition, we propose the pseudo class generation and metric learning method to enhance fine-grained recognition for retail products such as classification for products with multiple flavors. We demonstrate the effectiveness of each part of technique in our experiments for our target task, and show that our framework leads to high-accuracy recognition.