w

Abstract

Px.

Introduction

Fashion is a visual medium that is forever seeking out new channels to reach out to it's clientele. With the rapid proliferation of online media through blogs, social networks, e-commerce the field of fashion now has a landscape like never before. For a fashion e-commerce company, being able to filter such content and cater to its consumer segment is a challenging prospect. Fashion products are ephemeral and hence it becomes highly important for a fashion retailer to have a fresh and well curated catalogue. E-commerce retailers look at different media, like other online portals, fashion portals, AI generated designs etc. , for inspiration and identifying gaps in their catalogue.Further to identifying designs a retailer also needs to rank these new designs respective to their platform. However  a model learnt to gauge designs on a given platform from images might not work well for designs obtained from other media. This is because the underlying data distributions of the designs (images) on the 2 platforms could be very different. Intuitively this difference arises because styles are catalogued differently across mediums with differences in lighting , model poses , background etc.  AI generated images created through GANs, VAE etc., in practice lacks a faithful representation of training data and the generated images end up having a different distribution.
The generalisation capacity of any learning algorithm depends the fact that the test data follows the same underlying probability distribution as Training data used to learn the model. However there are many case where we fails to adhere to this axiom. In this work we develop a novel unsupervised domain adaptation algorithm to bring designs from different platforms to a common ground and learn a single model to grade styles.
In the literature, domain adaptation is solved in either a fully supervised framework or a semi-supervised framework. When we have huge annotated data from a given domain but we have little data in the domain of interest.  We want to take advantage of this situation to learn a deep learned model which performs well on  domain of interest. This scenario falls under the fully supervised framework. If the domain of interest is not annotated in any way but we can create some clusters to logically group and then learn a deep learned model which performs well on  domain of interest. This scenario falls under the semi-supervised framework. In our paper we use semi supervised framework.

Related WorkUnsupervised Domain AdaptationExperimentsConclusion

In this work we have attempted to address the Sales Potential of a fashion product in terms of its visual aesthetics. We have provided a formulation for estimating SP of a product by normalizing for the merchandising values and brand which introduces an inherent bias in its sales on the system. The above normalizing coefficients affecting SP of a style is then solved using Image similarity constraints (P-SP). We have shown that P-SP and DL-SP provide better SP score consistency on similar image pairs than a DP-SP retail model which is agnostic to visual aspects. We then show that P-SP score provides a significant correlation with CTR on the platform which captures some visual aspects of the product. We also model the SP by regressing with image features (DL-SP). We show that the P-SP and DL-SP approaches significantly outperform standard depricing model DP-SP (baseline) in terms of CTR correlations on image pairs. The analysis also shows that VGG-16 visual features may not capture the complete context of a style . Further, imposing similarity constraints in DL-SP formulation makes the model complex as the difference of image features may not be a good measure of similarity. Also the visual features needs a better articulation of the product which need to be learnt along with other context like trends, demography, customer segments etc. which we leave for future work. We finally show how the SP score is used in grading products in other e-commerce/social platform and for replenishment within the platform enabling better assortment planning.