Self-supervised Contrastive learning for Zero-shot automatic retail checkout
Annesha Nath · 2022 Innovations in Intelligent Systems and Applications Conference (ASYU) · 2022
Long queues at retail stores are a nightmare for every customer, and the present time-consuming checkout process degrades the overall shopping experience. In this paper, we propose an approach that can leverage the power of Deep Learning and Computer Vision to remove the time-taking human supervision from retail checkout systems. Specifically, we develop a Self-supervised Siamese network that takes the query image and original product image as input and verifies if the product is present in the query image. Previous approaches to developing automatic retail checkout systems in literature required data labeling and re-training the models every time a new product gets added to the retail database. In contrast, our model isn't trained on any labeled data, yet we show that our model generalizes to unseen products during the zero-shot evaluation phase, making it useful for real-world retail stores.