En | Ru

Improved photo search system for the marketplace

Client

KazanExpress - a trading platform with goods from Russia and China. Offers one day delivery

Background

01/

In 2020, EORA implemented its own MAGE image similarity search system on the KazanExpress website.

The customer decided to increase the accuracy of the search and gave us a new database of images: from stores and from users.

The system speeds up the purchase of goods: instead of using filters, the buyer simply uploads a photo.

Inside the service is a neural network that is able to analyze and find similar products.

Solution

Task

Post training of the neural network on new photos

Increase the quality of the search for similar images

Speed up the service amid growing number of requests

Optimization of service performance

How it works

02/

Buyer uploads a photo to search

The neural network accesses the database of the website and selects similar images

The customer sees a list of products with price tags and descriptions

Conversion from browsing to buying increases

Quote

03/

Alexey Guchko

«We did the project fast because we worked with an off-the-shelf product we know well: MAGE. We improved the search quality and also created a post-training module. It allows data to be indexed every time the image database is updated.»

Project Manager

Development Stages

04/

API preparation to launch the model

Step 03

Training a neural network on new photos

Step 02

Data markup: selection of relevant images

Step 01

Numbers

05/

minimum throughput of our computer model when running on a server

10 queries per second

X2 times

This is how much the search accuracy has increased after post training the model

Project challenges

Problem

Solution

In the new database, there were many not relevant photos received from users. The search may have included a photo of the packaging, but not of the product itself. Such images were not suitable for training.

We conducted several experiments and selected the training method that gives the highest search accuracy. To post train and test the computer model, we used photos of the seller (store).

06/

Development timeline

07/

15 days

September 2021

Interesting fact

Kazan Express has its own data science team. It creates all the product features - with the exception of tools based on computer vision technology. Those are handled by EORA