System for Deduplication of Machine Generated Designs from Fashion Catalog

Rajdeep H. Banerjee, Anoop Kolar Rajagopal, Vikram Garg, Sumit Borar

Abstract: A crucial step in generating synthetic designs using machine learning algorithms involves filtering out designs based on photographs already present in the catalogue. Fashion photographs on online media are imaged under diverse settings in terms of backgrounds, lighting conditions, ambience, model shoots etc. resulting in varying image distribution across domains. Deduping designs across these distributions require moving image from one domain to another. In this work, we propose an unsupervised domain adaptation method to address the problem of image dedup on an e-commerce platform. We present a deep learning architecture to embed data from two different domains without label information to a common feature space using auto-encoders. Simultaneously an adversarial loss is incorporated to ensure that the learned encoded feature space of these two domains are indistinguishable. We compare our approach with baseline calculated with VGG features and state of art CORAL [19] approach. We show with experiments that features learned with proposed approach generalizes better in terms of retrieval performance and visual similarity.