Neural networks · Computer vision · Fashion tech

Visual Content Analysis
for E-commerce & Advertisement Markets

AI tools for sales growth. A high-performance platform based on neural networks, commercially used in several massive scenarios:

Video-content-based ads targeting

Contextual advertising inside video streams, in real time

Fashion recommendations in e-commerce

Visual search, styling and try-on for online fashion shops

Video Content Targeting Engine

Makes video a showcase of goods and services — content-based advertising as a way of additional video monetization in the era of extended privacy protection standards.

Contextual advertising in video

As a general case

Fashion recommendations from video

As a special case

Features

Accurate and fast (in real time) visual content understanding, fashion recognition and fashion retrieval
Configurable and flexible framework with support of hardware-based scalability
Works with GPUs of different types
Distributed computing on servers and GPUs
Ready-to-go pipeline
API for simple integration

Benefits

A new sales channel through the ability to handle impulsive purchases
Greatly increases the size of advertising properties
Improves engagement and retention by making ads less intrusive (compared to mid-rolls and banners with behavior-based targeting)
Targeting possibilities in non-personalized channels (e.g. linear TV) via high-performance, real-time, automated, cost-effective labelling
Restores ads efficiency in the world of new privacy standards on the Internet
Enables a whole new array of services based on intelligent convergence between e-comm and media

Trusted

The engine was used to search for analogues of clothes from TV shows (in a third-party app) and for contextual ads on HbbTV.

1a · General case

Contextual Advertising in Video

Match video content with relevant ads and offers.

How it works

Object detection
Video analysis
Scene recognition
Video fragments classification
Brand Safety
Matching video fragments with categories of goods and services to place relevant ads

Special features

The most common types of scenes in typical video content are recognized and associated with ad categories — for example, food delivery, sporting goods, pets keeping and café services
High-accuracy scene type recognition

Table scene recognized in a video with a matching food delivery ad

1b · Special case

Fashion Recommendations from Video

Search for products similar to clothes from the video.

How it works

Object detection and tracking
Person appearance matching with one of the video characters
Gender recognition
Pose estimation
Clothes recognition
Features extraction for garments
Similar clothing items search by video, based on features similarity

Special features

A variety of supported object types: clothes, shoes, hats, bags and accessories such as ties and glasses
Works with millions of SKUs across multiple shops, keeping the catalog up to date on a daily basis
High-accuracy recommendations of similar items

Similar garments found in a catalog based on clothing detected in a video

Solutions for Fashion e-Commerce

Increase sales through efficient tools based on product appearance analysis.

AI-generated product tags and titles for a dress and boots

Automatic tagging & titles generation

AI tagging is an efficient alternative to manual tagging: the number of tags per product is doubled with similar accuracy.

Thousands of tags supported: category and model of clothing, shades of colors, texture, trim, material properties, style, gender, and special attributes such as sleeve shape and heel height.

Meaningful titles in a readable form.

Recognized color shades matched with clothing photos

Color shades recognition

140 color shades are recognized and can be used for filters and textual search.

Fine-grained color understanding turns “just red” into a precise, searchable attribute — Marsala, Wine, Scarlet or Turquoise.

Search for similar clothes

Similar offers search by image. A product photo — or any image or video uploaded by the user — can be used as a search sample.

This efficient alternative to text search provides smart navigation in online fashion shops.

All detected garments in the uploaded photo can be used as inputs for retrieval, enabling the user to recreate the whole outfit.

Outfitting recommendations

Complementary items search by image of reference clothes.

Generated combinations are perceived as advice from a human stylist.

Scalable solution: no manual labeling required. Cross-selling as a benefit.

Virtual Try-On

The generated image roughly predicts how the user would look in the selected dress. Two images required: a photo of the desired dress, demonstrated by a model in an online shop, and the user's photo.

VTO gives customers the opportunity to see their image in the selected garments and make informed decisions about purchase — removing barriers to online shopping and reducing returns.