LUN is the biggest accommodation search engine in Ukraine that processes 6 million ads per day, indexes 10 thousand websites and accepts 100 million visitors per year. How does the company that supports innovative approaches to trouble shooting use machine learning?
The head of AI team at LUN Volodymyr Kubytskyi answered this question in his lecture at Lunoteka co-working space.
AI as a marketing tool
On April 1, 2018, LUN website offered a new AI filter “With a carpet on the wall”. The algorithm was available for one week and over that time was in the spotlight of Ukrainian and foreign media.
Volodymyr states that the idea was born from the desire to emphasize the quality of renovation. Two neural networks were used to localize a carpet on the photo: the Region Proposal Network to generate candidate regions and the Object Detection Network.
“We correctly detect 99% of carpets on walls and our servers manage to process around 50 images per second ensuring constant operation of the filter,” states Andriy Mima,Co-founder & CFO of LUN.
This is how LUN’s AI filter for carpets looked like, DOU
“You buy Facebook, Google ads and measure brand awareness. At some point, when all channels get oversaturated, awareness stops growing. For example, when we measured awareness last time, we had 78% in Kyiv,” Volodymyr Kubytskyi says.
When a company is growing, brand awareness level is growing as well. To maintain the interest of users at a certain level, it is important to use several communication channels. At a certain point, the company decided to add outdoor advertising to the set of its tools.
Whereas everything is more or less clear with website metrics, how can one measure efficiency of outdoor ads?
“We wanted our ads to look good, unobtrusive, and did not want to compete with ads of other companies. We turned to an agency and said the following: we have one million hryvnia and we need billboards, advise us which ones to use. The agency told us: take these three – and that is all, we ran out of money. We asked why some billboards are very expensive and others are not,” Volodymyr shares, “It turned out that the agency used one quality criterion, namely, traffic volume. They did not take into consideration the season (for example, trees come into blossom and hide the ad), billboards of competitors, and complexity of traffic intersection. We wanted to spend our money more reasonably.”
Example of billboards, LUN
LUN’s AI development team decided to take charge of the assessment of outdoor ads. A comprehensive localization system was used to bring this idea to life. The system detects billboards in videos, calculates dynamic area in the field of view, calculates time within which a driver can look through the ad, and takes into account the competition with other billboards in real time.
“We drove an ordinary car and took video with an ordinary dashboard camera. We sent the video to the localization system applying the Kalman filter [keeps track of the estimated state of the system – editor's note] and wrapped it in the post processor,” the specialist explains, “In such a way, any car with any camera can find a billboard, gives it an ID, calculates the area it takes, and ads of rivals. Everything worked out.”
How comprehensive localization system works, LUN
The team has already tested the product in California. The trial went successfully, Volodymyr said.
From 2D to 3D using machine learning
In 2018, Facebook launched a feature that allowed posting 3D photos. To use the feature, one has to upload photos from the phone with two cameras.
This simplified algorithm enables shooting real three-dimensional films using a special movie camera with two lenses. Each of them makes shots for different eyes: the stereo effect relies on the fact that the right and left eyes see different shots.
Nevertheless, currently three-dimensional image can be created from two-dimensional, for example, the Titanic was filmed in 2D and converted in 3D in 15 years.
LUN decided to use the same solution to add interactive communication with content.
“We thought what if we cut 3D films into shots and taught the neural network to draw the image seen by the right eye basing on the image seen by the left eye? Knowing the focal distance and camera architecture that we imitate, we can calculate the depth map,” the lecturer explains.
In such a way, by learning on three-dimensional images, the neural network understands how flats are built and can create 3D images from ordinary photos.
Algorithm that enhances the online recommendations system
Usually users of accommodation search engines (especially those that look for flats to rent) are restricted in time – all lucrative offers quickly find their renters. And apart from the price and location, people want to choose an appropriate interior. To help users, LUN’s developers use machine learning.
“People have different tastes and often they search for flats basing on their preferences rather than some parameters. We have developed a system that can break the interior into components basing on photos and will try to find similar photos,” Volodymyr states.
Ads filtration, LUN
First the problem, then the solution
It is rather difficult to assess the economic benefit that machine learning brings to the company.
“To understand whether the algorithmic solution or a chatbot will bring value to the company, one has to ask one question: what problem I am trying to solve and whether this problem exists. We often want to develop something just because we want to. But we are always trying to find the problem first and then to solve it,” LUN’s Head of AI concludes.
Many new things about the efficient business use of machine learning
will be highlighted at AI Conference Kyiv held on June 4.
You'll know more than your colleagues and business rivals.