Neural network image search for ReRooms

Computer vision
System of search which is based on computer vision
Solution
Task
Ability to search for images of interior elements (furniture, accessories, floor coverings) in the ReRooms database. Interior designers work with this base.
Client
ReRooms is an online furniture store, finishing materials and household goods where you can also order a design project and turnkey repairs.
What does it look like from the inside
01/
Dataset of images which based on the material provided by client.
System of search based on MAGE platform
API for sending data of updates

Development stages

02/
Preparing data: indexing and describing images, deleting duplicates
Stage 1
Testing computer model the other
part of the images
Stage 4
Educating a neural network on one part of the images
Stage 3
API development for fast transferring database updates to the system
Stage 5
Selection of computer models, construction of hypotheses
Stage 2
Examples of how the system works
03/

Period of development

04/
1 month
July-August 2021
When we copy the client's database, ReRooms updated and we received extra part of data which cannot be deleted
Problem
All dataset of images had to be downloaded again. To do it we have to wait sunday's night, when ReRooms does not change
Solution
We have improved the API so that the search system quickly responds to changes in the database and re-indexes old and new products.
Solution
The assortment of the client's goods changes daily: for example, some sofas were sold out, others were brought.
Problem
Difficulties of project
05/

Findings

06/
When we work with dataset of images we found a lot of duplicates.
Due to this we avoid a lot of errors during the exploitation, at the same time client actualized info about their own range
Fact
The accuracy of the search of images is very high
Goods may have a different SKUs, name and description, photos may be taken from different angles. But the system quickly
determines the similarity.

Team

07/
Project Manager
Alexey Guchko
Data scientist
Vyacheslav Shults
Timlid
Vladislav Vinogradov
Ivan Izmailov
Backend-developer

Technologies

In development, we used our own solutions:
Integrated product search system by photo

Mage

EORA Neural Network Hosting Platform

Vision Hub

zDialog фреймворк
OneDash сервис
08/

Prospects

09/
Implementation of a recommender system on the ReRooms website (“They also buy with this product”, “You may also like” block).
By offering the user similar or related products, it increases the conversion of views into sales.
Recognition of goods on a separate area of the photo.
For example, a client is interested in a lamp - he circles it with the cursor, and the system selects similar products in the ReRooms catalog.

Similar projects

10/
Visual search of similar trade marks «Gardium»
INTELSONLINE – visual search of similar trade marks
KazanExpress – visual search of goods based on images
Contact us