Most websites do not tag attributes such as Occasion or Style in their catalogs, discovery becomes a challenge
We have Classifiers for Style for all product categories both Men and Women with accuracies of 75-80%
All product attribution is done manually
Can we train ML models to recognize the aesthetics of a garment- specifically Style
When thinking about fashion, customers don’t think in terms of product categories, they have a context in mind, such as Occasion or Style. However, most websites do not tag these attributes in their catalogs, so discovery becomes a challenge.
Imagine if you could ask the search engine” Show me a Boho dress that I could wear for Brunch” and it gave you all the relevant options. Wouldn’t that be fantastic?
Product search is the most critical touchpoint during a customer’s shopping journey. In order to enhance this process of product discovery, each product needs to be annotated with multiple layers of attributes to match the customer’s descriptions.
All product attribution is done manually. Some style-conscious sites spend the extra effort in attributing the product to its style. But for most brands, this is prohibitively expensive and they just go with the mandatory basic attributions.
HOW AI STUDIO HAS HELPED SOLVE IT
With the help of AI Studio, we used AutoML models to train Category classifiers to recognize the style
of a garment. The first step was to create a Taxonomy for Styles that were distinct and identifiable. - Boho, Glamorous, Classy, Relaxed, Safari, Retro, Sexy, Exotic, etc were the various values under the attribute Styles. It is imperative to ensure that these are discrete data sets else, the accuracy will be poor as the system gets confused with the similarities in the images.
Having clearly defined the style, the team collected data to ensure that we had all variations covered. We curated 80% eCommerce images some with plain backgrounds and some that were crowdsourced and not professional photographs. We also added 20% of street images with noisy backgrounds. The data bias towards eCommerce images was intentional as that was the final use case. The quality of training data directly impacts the accuracy of the model. Trying to recognize something so complex as a style that is open to interpretation was especially challenging. We defined Streamoid’s point of view and did a QC with that as a touchstone.
Once the data was collected for every attribute and label, we split the data set into training and test data. Each category was trained separately. For instance, Women's Topwear has its own set of models.
As this is an iterative process, the experiments are tracked in AI studio.
The accuracy for this particular style classifier is over 88%.
Some insightful suggestions from the model include:
1. Duplicate images are automatically removed, reducing clutter.
2. Confusion matrix counts between values.
3. Insights on failed images and the wrongly predicted value. This can be checked and corrected manually, thereby improving accuracy.
We have Classifiers for Style for all product categories both Men and Women with accuracies of 75-80% that can further be optimized for a specific dataset. Aesthetic attributes are a part of the data enrichment that we offer our clients for Cataloguing. It is also a critical attribute for our AI styling engine that creates theme-based curations and complete outfits.
Contact us to book a free demo.
MODEL BUILDER: Shruthi – Fashion Analyst and Classifiers Team