2. The goal & Problem
Choose proper product ID
from a product base by a
photo. Avoid a human
factor or lack of staff
experience.
For example, only a manager with extensive
experience can recognize broken remote
control with only half of the buttons present. A
recently hired employee cannot cope with
such a task — he needs to be trained for a
long time.
They all are very similar!
3. Visual search as a trend
Visual search is a growing trend: just take a
picture of what is interesting and the system
will find a product among millions of options.
You no longer need to describe the color, style
or other features of a product
Visual search is used by such global retail
companies as: ASOS, eBay, Neimann Marcus, Ted
Baker, Blippar, EasyJet, Levi’s, Disney, Walmart,
Salesforce, Syte, Houseology CarStory,
Snapchat’s, Farfetch, Marks, Pinterest, Amazon
4.
5. Our Solution
We spent near a year to find the best solution, that uses
custom neural networks, GPU-based servers to achieve
speed and accuracy
6. Visual
With the CNN ensemble-based solution
Product
Recognition
We’ve created
7. Preparing data
We prepared thousands of images of all the remotes,
with different backgrounds, lighting, and positions.
We mark each remote control picture, so the neural
network can understand where there’s an object the
systems needs to identify. We’ve created special
virtual studio, which generates markup images
(material) for neural network training. And we've
combined a dataset from this studio with manually
collected images
Our neural network made
over 6 million of steps
8. Solution features
As a result, our solutions can recognize a
remote from a photo1
too bright or dark
in hands
of erased buttons or labels
with a complex background
taken from an angle
9. Solution features
2 A neural network is able to
recognize remotes that look
identical. Dozens of remotes.
Hundreds.
Even the most experienced
employee, on whose training you
would spend a lot of time and funds,
could not show such accuracy and
speed of recognition
10. Solution features
The assembly of neural networks for the task of
recognizing remotes allows you to accurately determine
the model, and it does not depend on the language of
the buttons and labels
[IN THE ROADMAP] Solution will allow a re-training
without the participation of a developer. You will quickly
and efficiently train the network with new products
3
4
11. Comparison with hash-based
Hash-based algorithms are
used in Google’s visual search
and can quickly distinguish a
“cat” from a “car”, but not
more
Like Google visual search
cat car
cat same cat?
12. Comparison with SIFT/SURF
Algorithms that allow to detect
points of interest on an image,
but are too sensitive to light,
damage of an original object,
etc.
SIFT/SURF
Colosseum
Colosseum ???
Colosseum
13. Comparison with custom CNN
(as in Google Lens, Google
AutoML Vision and in other
boxed solutions) - allows you
to perfectly distinguish one
class of the remote from
another, but makes mistakes if
remotes are visually similar
Google AutoML,
Google Lens
Different
Same?
14. Technologies used
TensorFlow
The machine learning framework. We used it create the neural network with
the model that is optimally suitable for further training
Google Cloud Vision, Soundex
To recognize labels, we used the Google Cloud Vision API, and to find the
best match among the possible results, we used full-text search, Levenshtein
distance calculation, and Soundex. Soundex is an algorithm for comparing
two strings by their sound, setting the same index for strings that have a
similar sound in English.
15. Solution business values
Visual search is a growing
trend in retail. Book your
place now!
Help wholesale partner.
Save their costs for staff
training and give them a
trendy tool for
end-customers
Be an innovative
company!
Unique solution for your
company
Easy integration
with API service
Algorithms are better than
used in analogues
16. Business value
Services for an end-customer:
- Integrate it to chatbots, websites, mobile apps of
your wholesales partners and allow your
end-customers to find a remote by photo
- Achieve an advantage: you will be on a priority list
because your goods are easier to find
- Measure a demand: collect what models appends
in search often than others
17. Future development
A system that allows upload photos
of a new product and retrain a neural network (current network
is trained on 90 products from several thousand of remotes)
Mobile application + server for the automated markup of training
material (now a lot of manual labor is used to prepare the
material for training, we can automate it)
1
18. Future development
A control panel with a dashboard that allows you to understand
how many times a particular remote has been searched.
Notifications about unrecognized items and re-training with them
2
19. Pricing model
The client gets an access to the API service and the control panel to view a statistics
Yearly
Per
scan
Delivery of the solution
as a ready-made boxed solution
with an annual subscription
Payment for each scan /
package scans
or