Five years ago, the ways visual artificial intelligence (AI) was being applied in fashion e-commerce was narrow and limited. Now its tentacles spread far and wide – from achieving back-end efficacy to improving front-end search results. Visual AI is showing strong growth potential in the retail segment with 30 percent of retailers adopting some form of visual AI technology. This research brief explores some of the primary applications of visual AI in fashion e-commerce.
1. AI-assisted Product Tagging for Cost-effective and Efficient Catalog Management
Manual coding of products is expensive, non-scalable and slow, with high chances of error. Retailers take 25 minutes per SKU to manually update product data, which translates to a 20-30 day timeframe for a product to go online. Manual product tagging is untenable, given the industry’s scale, the pace of new product introductions, compressed fashion cycles and rising expectations of a flawless customer experience (CX). Moreover, inadequate and inexact tagging leads to irrelevant search results and product recommendations, and inaccurate categorization in catalogs, which can lead to as much as a 30 percent loss in purchase potential.
Today, advanced AI technologies, such as convolutional neural networks (CNNs), can extract attributes such as color, category, style, pattern, cut, shape, length, size and brand from fashion images. Fashion retailers are eyeing these tools for their potential value in e-commerce.
An AI system for automated tagging uses two primary technologies – computer vision and a deep learning algorithm to create what is essentially a neural network that can mimic the human brain in identifying different attributes of an image. For AI to work effectively in this context, it needs to analyze thousands of images to understand how a piece of clothing looks and the type of tags that can best describe it. The computer vision identifies the image, and the deep learning algorithm helps process the pixel content of visuals, discover relevant objects and extract the required characteristics. After a period, the deep learning algorithm also can be programmed for additional tasks such as determining the connection between tags, which tags are being used and any patterns in usage that might exist.
The primary benefit for retailers is more cost-effective (operational cost can reduce up to 90 percent) and faster (time can be reduced up to 90 percent) catalog management, along with rich and consistent metadata for improved product discovery.
Some of the other topline and CX benefits that make visual AI a compelling proposition include:
- Improved SEO: As 35 percent of e-commerce traffic comes from organic search, an updated and accurate product catalog can boost SEO performance.
- Enhanced conversion, average order value and CX: Because visual AI can improve the recommendation engine, the cart abandonment rate goes down (reducing null results by as much as 10 percent), leading to increased basket size. Vendors operating in the segment are seeing an increase in conversion by 12-30 percent, with an increase in order value of up to 10 percent.
- Better personalization: If products have a large number of meaningful attributes (which is closer to how consumers would describe a product), every algorithm that shapes the purchase path works better.
- Reduced time to market: This is ensured through faster onboarding of new SKUs.
- Optimized purchase path: Comprehensive and accurate tagging improves organic search results and enhances the chances of product discovery. Concurrently, more diverse landing pages can correspond better with search strings and results.
- Improved inventory management: AI-assisted attribute tagging software uses a large lexicon of widely adopted tags. This leads to regional and segment-specific product tags for different customer segments, customer tiers and geographies, reinforcing greater discipline in inventory management.
Tech-savvy retailers such as ASOS have developed in-house machine learning models to accomplish automated product tagging. Other retailers are relying on independent software vendors (ISVs), including Catchoom, Vue.ai, Pixyle, Syte, Wide Eyes Technologies, Visenze, Fabulyst, Trendage, Intelistyle.
2. Visual Search for Instant Gratification
Three trends come together to make a strong business case for visual search: 1) the impulsive buying patterns of Gen Z (which forms almost 40 percent of e-commerce shoppers), 2) use of social networks such as Snapchat and Instagram for product placement, and 3) the inherent human ability to process visuals 60,000 times faster than text. Visual search uses the principles and technologies also used by AI systems for automated tagging, but with one primary difference ― after decoding the product attributes of an image, the algorithm then searches, retrieves and displays images (with matching attributes) to the shopper.
Enhanced customer experience through the quick and accurate discovery of products remains the primary value proposition of visual search – more than 60 percent of millennials and Gen Z deem visual search capability as more important than any other new technology when it comes to the retail customer experience.
Other related benefits that can be achieved through visual search include:
- Similar-items recommendations: A visual search not only jumpstarts product discovery but provides access to visually similar items, allowing greater product exposure. Shoppers can compare prices, delivery dates and user-generated buyer tips that, together, ensure a more holistic shopping experience.
- Increased conversion rate: Consumers using visual search are essentially in a purchase-ready mode, therefore, the chances of a completed transaction are higher.
- Better inventory management and digital merchandising: Visual search results can generate insights into products that are experiencing greater interaction and purchase volume. This, in turn, can help in planning new inventory and determining the effectiveness of inventory display on a homepage.
3. Automated Human Model Imagery to Save Cost and Increase Conversion
Typically, increasing conversion and average order value rates in fashion e-commerce requires professionally shot images of a model displaying the product. The entire process is expensive and non-scalable.
In 2014, a new technology emerged in the domain of computer vision and image processing when Ian Goodfellow, a machine learning research scientist, invented generative adversarial networks (GANs). At the same time, a related and relatively older technology CNNs saw a revival. CNNs allow the capture of higher-order image attributes. Using generative technology and CNN, it is possible to capture a pose (such as a head tilt) from a source image and apply it to a digitally generated image.
The use of GANs and CNNs has the potential to change product/apparel display in fashion e-commerce. An apparel can be added digitally to the stock image of a model. Also, a stock model image can be manipulated digitally (and automatically) to alter the pose, mood or ensemble for a brand. Adding new images to the same SKU is easy with the use of AI. Costs do not scale linearly, allowing the use of multiple models of different body types or ethnicities. The result is an e-commerce experience that is more highly personalized, allowing greater resonance with a brand.
Reduced cost of photography remains the primary benefit of the technology, while other benefits include:
- On-model images can reduce the return rate by 6.5 percent: Digitally rendered catalog models of various body types have shown to be cost-effective. Current solutions such as TrueFit offer shoppers a “virtual try-on” experience, but it requires conscious recall and extra clicks. In comparison, the size-specific model approach is intuitive, does not slow the customer down or allow a second guess on the purchase. It is a more seamless purchase path, where a user would not have to spend time remembering (or physically checking) size for a particular brand.
- Better personalization: With generative algorithms, retailers can economically display models with varying skin tone, body type and overall “look” aligned to customers, thereby achieving better personalization.
- Possible individualization: Using generative models to transfer style and pose elements from a source to a target, it would be possible to drape apparel (in 3D) on a shopper’s own image and gauge a “look” from a high-quality e-commerce site.
- Better ensembles and styling: Using deep neural networks, retailers can identify articles of clothing that are stylistically related and/or co-occur frequently. This also helps maintain cost margins as digitally adding outfits to a model does not involve photoshoot time or costs.
- More granular SKUs: When retailers add permutations such as color to a product type, the SKU count explodes—85,000 SKUs can expand to 500,000. Since the marginal cost of adding a new color or some other minor variant is near zero with AI, the e-commerce shopper can get much closer to how a product would actually look.
Currently, adoption of this technology is in a nascent stage; only a handful of tech firms like Vue.ai and Flixstock offer AI-based human model imagery. However, given its value proposition, retailers can anticipate an influx of new technology firms offering automated human model imagery.
ISG understands the quickly changing technology landscape. We work with retailers to find the right technology partners to help improve customer experience, reduce operational costs and get ahead of the competition. Contact us to get started.